fix: audit batches 3–4 — TLS verify, WS, OFFSET, B-tree, NULL equality

Close the remaining 2026-08 findings: peer TLS on leader forward, disttxn
SO_ERROR, compaction catalog order, OFFSET without LIMIT, window aggregates,
WebSocket mask/size/auth, SCRAM timing and cbind, B-tree leaf left-max
separators, and SQL three-valued NULL comparisons.

.github/workflows/ci.yml

@@ -61,6 +61,30 @@ jobs:
           done
           echo "--- Done ---"
 
+  raft-e2e:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - name: Setup Nim
+        uses: jiro4989/setup-nim-action@v1
+        with:
+          nim-version: '2.2.10'
+      - name: Install system dependencies
+        run: sudo apt-get update -qq && sudo apt-get install -y -qq libssl-dev libpcre3-dev openssl ca-certificates
+      - name: Install Nim dependencies
+        run: nimble install --depsOnly -y
+      - name: Build server
+        run: nim c -d:ssl -o:build/baradadb src/baradadb.nim
+      - name: Raft e2e suites
+        env:
+          CI: "true"
+        run: |
+          nim c -d:ssl --threads:on --path:src -r tests/raft_e2e_test.nim
+          nim c -d:ssl --threads:on --path:src -r tests/raft_writes_e2e_test.nim
+          nim c -d:ssl --threads:on --path:src -r tests/raft_failover_load_e2e_test.nim
+          nim c -d:ssl --threads:on --path:src -r tests/raft_tls_e2e_test.nim
+          nim c -d:ssl --threads:on --path:src -r tests/raft_coldnode_e2e_test.nim
+
   verify:
     runs-on: ubuntu-latest
     steps:

.gitignore

@@ -16,6 +16,9 @@ tests/bugfix_test
 tests/nimforum_smoke_test
 tests/raft_e2e_test
 tests/raft_writes_e2e_test
+tests/raft_failover_load_e2e_test
+tests/raft_tls_e2e_test
+tests/raft_coldnode_e2e_test
 benchmarks/bench_all
 benchmarks/compare
 clients/nim/tests/test_client
@@ -68,5 +71,6 @@ src/barabadb/storage/lsm
 src/barabadb/storage/wal
 src/barabadb/storage/btree
 src/barabadb/storage/gate
+src/barabadb/protocol/scram
 clients/nim/tests/test_pool
 clients/nim/tests/test_wire

BUG_AUDIT_2026-08.md

@@ -0,0 +1,80 @@
+# BaraDB — Deep Audit (август 2026)
+
+> Дата: 2026-08-02
+> Метод: 4 паралелни одит-агента по слоеве (Storage / Query / Core / Protocol), всеки чете всички файлове в обхвата си и проверява находките срещу реалния код.
+> Обхват: **само нови дефекти** — 80-те вече оправени в `BUGS.md` / `BUG_AUDIT.md` / `BARADB_CLIENT_BUGS.md` са изключени.
+> **Общо: ~28 находки | Поправени: 28 (батч 1: 5 + батч 2: 12 + батч 3: 10 + батч 4: 2) | Остават: 0**
+
+---
+
+## Поправени — батч 1 (5)
+
+| # | Severity | Проблем | Файл | Fix |
+|---|----------|---------|------|-----|
+| C1 | 🔴 CRITICAL | **MIGRATE handler без auth gate** — неавтентикиран клиент пишеше произволни key/value в базата (`handleMigrationMessage → applyMigrationBatch → storeKeys → db.put`). Открито независимо от 2 агента. | `core/server.nim:682` | Добавен `if not authenticated: ... continue` (като REP/DISTTXN блоковете) |
+| C2 | 🔴 CRITICAL | **Raft commit quorum off-by-one за even-N** — `(N+1) div 2` commit-ваше с малцинство при четен брой възли (N=4 → 2/4). Election-ът ползваше коректното strict majority. *GA обхватът е 3-node (нечетно), където формулите съвпадат.* | `core/raft.nim:653` | `let majority = (node.peers.len + 1) div 2 + 1` (съвпада с election); регресионен тест за 4-node |
+| H1 | 🟠 HIGH | **Pre-auth memory-exhaustion DoS** — `parseHeader` не ограничаваше `length` (uint32 до ~4 GiB); `recvExactWithTimeout` пре-алокира преди auth check. | `core/server.nim:168` | Reject `length > uint32(MaxWireStringLen)` (64 MB) преди алокация |
+| H4 | 🟠 HIGH | **`**` и `++` се lower-ваха към equality** — `bkPow`/`bkConcat` липсваха в op-mapping case-а и попадаха в `else: irEq` (`2 ** 3` → `false`, `'a' ++ 'b'` → `false`). | `query/exec/lower.nim:79` | `of bkPow: irOp = irPow`, `of bkConcat: irOp = irAdd`; 2 регресионни теста |
+| H5 | 🟠 HIGH | **`!=` не е отрицание на `=`** — `irNeq` short-circuit-ваше на string inequality, така че `5 != 5.0` → true, но `5 = 5.0` → true. | `query/exec/eval.nim:438` | `irNeq` numeric-first (точно допълнение на `irEq`); регресионен тест |
+
+## Поправени — батч 2 (12)
+
+| # | Severity | Проблем | Файл | Fix |
+|---|----------|---------|------|-----|
+| H3 | 🟠 HIGH | **Semi-sync partial/zero ack** — връщаше LSN дори при 0 acks | `core/replication.nim` | `return 0` когато connected replicas < `syncReplicaCount` acks; 0 connected → local-only (като sync) |
+| H6 | 🟠 HIGH | **`COUNT/SUM/AVG(DISTINCT)` игнорира DISTINCT** | `query/exec/lower.nim`, `plan_exec.nim` | `aggDistinct = node.funcDistinct`; dedup с `HashSet` в agg пътищата |
+| H7 | 🟠 HIGH | **`UNION/INTERSECT/EXCEPT` KeyError** | `query/executor.nim` | Dedup fingerprint от projected cols, не `row["$value"]` |
+| H8 | 🟠 HIGH | **`MERGE … THEN DELETE` / matched condition no-op** | `query/executor.nim` | Honor `mergeMatchedDelete` + `mergeMatchedCondition` |
+| H9 | 🟠 HIGH | **WAL recovery crash на torn record** | `storage/lsm.nim`, `wal.nim`, `recovery.nim` | Bound key/val ≤ 64 MB; validate kind преди enum cast |
+| M1 | 🟡 MEDIUM | **MVCC `write` delete-during-iteration** | `core/mvcc.nim` | Collect-then-delete stale txn ids |
+| M3 | 🟡 MEDIUM | **`checkpoint` lock leak** | `storage/lsm.nim` | try/finally около write lock + walLock |
+| M4 | 🟡 MEDIUM | **`flushUnsafe` clear-before-write** | `storage/lsm.nim` | Clear memtable едва след успешен `writeSSTable` |
+| M5 | 🟡 MEDIUM | **Compaction empty-key skip** | `storage/compaction.nim` | `haveLast` флаг вместо `lastKey = ""` sentinel |
+| M6 | 🟡 MEDIUM | **`rewriteLive` remove-before-move** | `storage/wal.nim` | Само атомен `moveFile` (rename replace) |
+| L3 | 🟢 LOW | **mmap `offset+size` overflow** | `storage/mmap.nim` | Overflow-safe: `offset > size - length` |
+| — | hygiene | **Stray ELF `protocol/scram`** | `.gitignore` | Премахнат binary + ignore entry |
+
+**Верификация (батч 2):** `baradadb` build чист; `tests/bugfix_test.nim` (вкл. batch-2 suite) и `tests/test_all.nim` (501 OK) минават без `[FAILED]`. `tests/prop_test.nim` B-Tree suite OK (H10 *не* е в този батч — naive left-max fix чупи interleaved remove).
+
+## Поправени — батч 3 (10)
+
+| # | Severity | Проблем | Файл | Fix |
+|---|----------|---------|------|-----|
+| H2 | 🟠 HIGH | **TLS client връзките не верифицираха сертификата** | `core/server.nim`, `core/config.nim` | Отделен `tlsClient` контекст; CA auto-enable verify; production fail-closed |
+| M2 | 🟡 MEDIUM | **disttxn refused-connect wedge** | `core/disttxn.nim` | `getsockopt(SO_ERROR)` + try/except около RPC |
+| M7 | 🟡 MEDIUM | **Compaction unlink преди catalog load** | `storage/compaction.nim`, `baradadb.nim` | Unlink след load + MANIFEST |
+| M8 | 🟡 MEDIUM | **`OFFSET` без `LIMIT` → 0 реда** | `query/exec/lower.nim`, `plan_exec.nim` | `limitCount = -1` unlimited; clamp negative |
+| M9 | 🟡 MEDIUM | **Window SUM/AVG/COUNT/MIN/MAX → NULL** | `query/exec/window.nim` | Frame aggregates |
+| M10 | 🟡 MEDIUM | **WebSocket unmasked client frames** | `core/websocket.nim` | Protocol error / close |
+| M11 | 🟡 MEDIUM | **WebSocket unbounded buffer** | `core/websocket.nim` | 1 MiB frame / 4 MiB message / 125-byte control |
+| M12 | 🟡 MEDIUM | **SUBSCRIBE без table auth** | `core/websocket.nim`, `httpserver.nim` | `canSubscribe` + `hasPrivilegeFor` SELECT |
+| L1 | 🟢 LOW | **SCRAM timing user enumeration** | `protocol/auth.nim` | Dummy nonce+encode work за unknown users |
+| L2 | 🟢 LOW | **SCRAM `c=` не се верифицира** | `protocol/auth.nim`, `scram.nim` | `c=` must match gs2 header (`biws` за `n,,`) |
+
+## Поправени — батч 4 (2)
+
+| # | Severity | Проблем | Файл | Fix |
+|---|----------|---------|------|-----|
+| H10 | 🟠 HIGH | **B-tree `remove` separator convention** — `splitChild`/search са left-max (`key > sep → right`); `removeRec` копираше first key на дясното дете (right-min). Naive left-max върху *internal* ключове чупи `prop_test`. | `storage/btree.nim` | Leaf borrow/remove пишат left-max; merge underflow се качва нагоре; invariant `max(left) <= sep` (boundary duplicates са позволени, `next` ги събира) |
+| L4 | 🟢 LOW | **NULL equality** — `NULL = NULL` / `col = NULL` бяха true през string sentinel | `query/exec/eval.nim` | Сравнения, LIKE, IN, NOT, AND/OR: NULL operand → unknown (`\N`); `IS NULL` непроменен |
+
+---
+
+## Остават (0)
+
+Няма отворени находки от този одит.
+
+---
+
+## Проверени и чисти (не са бъгове)
+
+- JWT `exp`/alg-confusion: pinned `jwt-nim-baraba#fbe084b` `verify()` enforce-ва alg-match, reject-ва `NONE`, проверява `exp/nbf/iat`, constant-time compare.
+- `auth.nim` `constantTimeCompare` и SCRAM `verifyClientProof` са constant-time; празен JWT secret fail-closed (`server.nim:59`).
+- `wire.nim` deserialize bounds/depth caps са sound.
+- CRC byte ranges / `headerSize = 40` са консистентни между write/verify/load (format *коментарът* още казва "36" — само коментар).
+- Lock ordering (`walLock` в `db.lock`; gate преди `db.lock`), mmap negative offset / `close()` recursion / fd handling (BUG-036/046) — непокътнати.
+- `COUNT(col)` изключва NULLs (`v.kind != vkNull`); `LIMIT 0` → празно е коректно; IN-list се lower-ва към OR/AND вериги.
+
+---
+
+*Виж също: `PLAN.md` (Сесия 13), `docs/en/known-limitations.md`.*

CHANGELOG.md

@@ -2,6 +2,78 @@
 
 All notable changes to BaraDB are documented in this file.
 
+## [Unreleased]
+
+### Security
+
+- **MIGRATE auth bypass (CRITICAL)** — the internal `MIGRATE` text-protocol handler now requires authentication (matching the `REP`/`DISTTXN` handlers); previously an unauthenticated client could inject arbitrary key/value rows (`core/server.nim`)
+- **Pre-auth wire-length DoS (HIGH)** — `parseHeader` rejects messages larger than the 64 MB wire cap before allocating the receive buffer (`core/server.nim`)
+- **TLS peer verify on cluster forwarding (HIGH)** — follower→leader SQL forwarding uses a dedicated client TLS context; `BARADB_TLS_CA_FILE` auto-enables `BARADB_TLS_VERIFY_PEER`; production fails closed if TLS is on without CA+verify (`core/config.nim`, `core/server.nim`)
+- **WebSocket unmasked frames / unbounded buffers (MEDIUM)** — unmasked client frames are rejected (RFC 6455 §5.1); frame/message size and control-frame caps bound memory (`core/websocket.nim`)
+- **WebSocket SUBSCRIBE table auth (MEDIUM)** — with auth enabled, SUBSCRIBE requires table SELECT privilege (`core/websocket.nim`, `core/httpserver.nim`)
+- **SCRAM timing + channel-binding (LOW)** — unknown users do equivalent dummy work; `c=` must match the gs2 header (`protocol/auth.nim`, `protocol/scram.nim`)
+
+### Fixed
+
+- **Raft commit quorum (CRITICAL)** — commit now requires a strict majority (`N div 2 + 1`), matching the election check; the previous `(N+1) div 2` formula committed at a minority for even-sized clusters (`core/raft.nim`)
+- **`**` / `++` operators (HIGH)** — power and concat are no longer lowered to equality: `2 ** 3` → 8, `'a' ++ 'b'` → `'ab'` (`query/exec/lower.nim`)
+- **`!=` semantics (HIGH)** — `!=` is now the exact complement of `=` for numerically-equal values (`5 != 5.0` is false) (`query/exec/eval.nim`)
+- **Semi-sync partial ack (HIGH)** — `writeLsn` in `rmSemiSync` returns `0` when connected replicas fail to meet `syncReplicaCount`; zero connected peers still succeed local-only (like sync) (`core/replication.nim`)
+- **`COUNT/SUM/AVG(DISTINCT …)` (HIGH)** — `funcDistinct` is copied to `aggDistinct` and applied via `HashSet` dedup in aggregate paths (`query/exec/lower.nim`, `plan_exec.nim`)
+- **`UNION` / `INTERSECT` / `EXCEPT` (HIGH)** — set-op dedup fingerprints projected columns instead of missing `row["$value"]` (KeyError crash) (`query/executor.nim`)
+- **`MERGE … WHEN MATCHED THEN DELETE` (HIGH)** — executor honors `mergeMatchedDelete` and optional `mergeMatchedCondition` (`query/executor.nim`)
+- **WAL recovery torn records (HIGH)** — recovery bounds key/value to 64 MB and rejects out-of-range entry kinds before enum cast (avoids multi-GiB alloc / `CaseStmtError` Defect) (`storage/lsm.nim`, `wal.nim`, `recovery.nim`)
+- **MVCC `write` timeout cleanup** — stale active transactions are collected then deleted (no mutation during `activeTxns` iteration) (`core/mvcc.nim`)
+- **`checkpoint` lock leak** — write lock and `walLock` released in `try/finally` (`storage/lsm.nim`)
+- **`flushUnsafe` data-loss window** — memtable is cleared only after a successful SSTable write (`storage/lsm.nim`)
+- **Compaction empty-string key** — dedup uses a `haveLast` flag so key `""` is not skipped (`storage/compaction.nim`)
+- **`rewriteLive` crash window** — atomic `moveFile` replace only (no `removeFile` before rename) (`storage/wal.nim`)
+- **mmap OOB on overflow** — length checks use `offset > size - length` instead of wrapping `offset + size` (`storage/mmap.nim`)
+- **REP replication put/delete encoding** — the legacy REP payload carries an explicit op tag so PK-only inserts (empty value) replicate as puts instead of vanishing as deletes (`core/replication.nim`, `core/server.nim`)
+- **REP receiver secondary indexes** — the legacy REP receiver applies via `applyReplicatedPut/Delete` under the storage gate, keeping B-tree/FTS/HNSW/graph indexes consistent on the replica (`core/server.nim`)
+- **Snapshot send stall (partial)** — the leader's snapshot send runs gzip off the event loop on a worker thread (`gzipFileAsync`), so heartbeats keep firing during compression; tar (send) and the restore path still run on the loop (`core/backup.nim`, `core/raft.nim`)
+- **disttxn refused-connect wedge (MEDIUM)** — `connectWithTimeout` checks `SO_ERROR`; `sendDistTxnRpc` catches `CatchableError` so a refused peer cannot leave 2PC stuck (`core/disttxn.nim`)
+- **Compaction catalog order (MEDIUM)** — inputs stay on disk until the output is loaded and MANIFEST is written; then they are unlinked (`storage/compaction.nim`, `baradadb.nim`)
+- **`OFFSET n` without `LIMIT` (MEDIUM)** — no longer treated as `LIMIT 0`; negative LIMIT/OFFSET are clamped (`query/exec/lower.nim`, `plan_exec.nim`)
+- **Window `SUM`/`AVG`/`COUNT`/`MIN`/`MAX` (MEDIUM)** — aggregate window functions compute over the frame instead of returning NULL (`query/exec/window.nim`)
+- **B-tree leaf separators (HIGH)** — leaf borrow/remove now keep the left-max convention used by `splitChild` and search (`key > sep → right`); merge underflow rebalances up the tree; `checkInvariants` guards `max(left) <= sep` (`storage/btree.nim`)
+- **NULL comparison three-valued logic (LOW)** — `NULL = NULL` / `col = NULL` / `col != x` yield unknown (`\N`) so `WHERE` excludes them; `IS NULL` is unchanged (`query/exec/eval.nim`)
+
+### Removed
+
+- Stray compiled ELF `src/barabadb/protocol/scram` from the source tree (added to `.gitignore`)
+
+### Added
+
+- Deep audit report `BUG_AUDIT_2026-08.md` (~28 findings; all fixed across batches 1–4)
+
+---
+
+## [1.3.0] — 2026-07-30
+
+### Raft cluster — Supported (single `default` DB scope)
+
+Raft 3-node moves from **Experimental** to **Supported** for the covered scope; single-node remains Production GA. Spec/plan: `docs/superpowers/specs/2026-07-30-raft-supported-design.md`, `docs/superpowers/plans/2026-07-30-v1.3.0-raft-supported.md`.
+
+- **Failover under load proven** — `tests/raft_failover_load_e2e_test.nim`: sustained INSERT load, leader killed at ≥ 50 acked writes; every **acknowledged** write survives on both survivors. Client contract: in-flight writes during failover fail fast with an error — clients must retry ([distributed.md](docs/en/distributed.md))
+- **Mandatory CI gate** — dedicated `raft-e2e` GitHub Actions job runs all five raft e2e suites; a missing server binary is a hard FAIL under CI (no silent skip)
+- **Raft-port TLS** — `BARADB_RAFT_TLS_ENABLED` + `BARADB_RAFT_TLS_CERT_FILE` / `BARADB_RAFT_TLS_KEY_FILE` / `BARADB_RAFT_TLS_CA_FILE` / `BARADB_RAFT_TLS_VERIFY_PEER`; fail-closed startup when cert/key is missing; optional mutual auth; follower→leader SQL forwarding is TLS-wrapped when the client port is. E2E `tests/raft_tls_e2e_test.nim` (full-TLS cluster works; plaintext node excluded)
+- **InstallSnapshot cold-node recovery** — backward-compatible wire protocol (`RaftProtoVersion` stays 1); the leader streams a `tar.gz` snapshot of the default DB in chunks of `BARADB_RAFT_SNAP_CHUNK_KB` KiB (default 256) to peers whose lag is unrecoverable; the follower restores via the backup/restore path and resumes from the snapshot base. Leader compaction unpins from peers stale beyond `BARADB_RAFT_PEER_STALE_MS` (default 30000). E2E `tests/raft_coldnode_e2e_test.nim` (returning node and wiped node converge automatically)
+
+### Fixes
+
+- **Raft put/delete encoding** — `ExecResult.keyValuePairs` carries an explicit `deleted` flag; an INSERT into a PK-only table (empty value) is no longer encoded as a `delete` and erased on apply; regression tests in `tests/bugfix_test.nim`
+- **Rejoin livelock** — on leadership acquisition the leader drops cached peer sockets; half-dead sockets to a restarted peer previously never errored, so no redial ever happened and the cluster livelocked without heartbeats
+- **Post-restore ctx repoint** — after an InstallSnapshot restore, the TCP serving ctx/db is repointed at the reopened database (queries previously read the closed pre-restore LSM and served 0 rows). Remaining limitation for startup-captured HTTP ctx: see [known-limitations](docs/en/known-limitations.md)
+- **Intermediate InstallSnapshot chunk replies ignored** — the leader acts only on the final chunk reply
+
+### Release
+
+- `baradadb.nimble` → `1.3.0`; `/health` and startup version strings updated
+- Docs: [distributed](docs/en/distributed.md) (failover contract, raft TLS setup, snapshot tunables), [known-limitations](docs/en/known-limitations.md) (raft supported scope + two newly documented limitations), [release-checklist](docs/en/release-checklist.md)
+
+---
+
 ## [1.2.0] — 2026-07-30
 
 ### Production GA (single-node)

PLAN.md

@@ -145,6 +145,34 @@
 
 ---
 
+## Сесия 13: Stabilization & Deep Audit (август 2026)
+
+> **Цел**: Подобряване на надеждността и коректността върху вече завършеното ядро — целеви поправки + системен паралелен одит по слоеве (Storage / Query / Core / Protocol).
+
+### Целеви поправки (завършени)
+
+| # | Поправка | Файлове | Статус |
+|---|----------|---------|--------|
+| 1 | **REP delete-from-empty** — legacy REP payload носи явен put/delete таг (`encodeRepPayload`/`decodeRepPayload`); PK-only редове вече не изчезват при репликация | `core/replication.nim`, `core/server.nim` | ✅ + тестове |
+| 2 | **Snapshot stall (частична mitigation)** — gzip при leader snapshot send се изнася извън event loop-а през worker thread (`gzipFileAsync`); heartbeats текат по време на компресия | `core/backup.nim`, `core/raft.nim`, `baradadb.nim` | ✅ + e2e |
+| 3 | **REP receiver индекси** — receiver-ът минава през `applyReplicatedPut/Delete` под storage gate; вторичните индекси (B-tree/FTS/HNSW/graph) се поддържат на репликата | `core/server.nim` | ✅ + тест |
+
+### Deep Audit (2026-08)
+
+4 паралелни одит-агента по слоеве; ~28 нови находки (без дублиране на 80-те вече оправени в `BUGS.md`/`BUG_AUDIT.md`). Пълен отчет: [`BUG_AUDIT_2026-08.md`](BUG_AUDIT_2026-08.md).
+
+**Батч 1 — поправени (5):** MIGRATE auth bypass (CRITICAL), raft commit strict-majority за even-N (CRITICAL), pre-auth wire-length DoS (HIGH), `**`/`++` lowering към equality (HIGH), `!=` не е отрицание на `=` (HIGH).
+
+**Батч 2 — поправени (12):** semi-sync partial-ack (H3), COUNT/SUM/AVG(DISTINCT) (H6), UNION/INTERSECT/EXCEPT (H7), MERGE THEN DELETE (H8), WAL torn-record recovery (H9), MVCC write iteration (M1), checkpoint lock leak (M3), flushUnsafe order (M4), compaction empty-key (M5), rewriteLive atomic replace (M6), mmap overflow (L3), stray `protocol/scram` ELF.
+
+**Батч 3 — поправени (10):** TLS peer verify (H2), disttxn SO_ERROR (M2), compaction catalog order (M7), OFFSET-без-LIMIT (M8), window агрегати (M9), WebSocket (M10–M12), SCRAM (L1–L2).
+
+**Батч 4 — поправени (2):** B-tree leaf left-max separators (H10), NULL three-valued comparisons (L4).
+
+**Остават (0):** вж. `BUG_AUDIT_2026-08.md`.
+
+---
+
 ## Какво остава от старите планове
 
 | Стар план | Статус |
@@ -157,6 +185,7 @@
 | **Този план** — Сесии 10, 11, 12 | ✅ Завършен |
 | Raft C3a/C3b + DDL/forward/compact/metrics (2026-07-30) | ✅ Завършен на `main` — `docs/superpowers/specs/2026-07-30-raft-cluster-status.md` |
 | **Production GA v1.2.0** (single-node) | ✅ `docs/superpowers/plans/2026-07-30-production-ga.md` |
+| **Сесия 13** — Stabilization & Deep Audit (2026-08) | ✅ Батч 1–4 (28 поправки); `BUG_AUDIT_2026-08.md` |
 
 ---
 
@@ -170,4 +199,4 @@
 
 ---
 
-*План версия: 2026-05-17*
+*План версия: 2026-08-02*

README.md

@@ -1576,7 +1576,7 @@ features are still being refined:
 | LSM-Tree SSTable reads | ✅ Implemented | Full disk I/O with compaction, WAL, and bloom filters. |
 | HNSW vector search | ✅ Implemented | Hierarchical graph navigation with SIMD-optimized distance metrics. |
 | TCP server execution | ✅ Implemented | Full binary wire protocol parsing and BaraQL query execution. |
-| Raft consensus | ⚡ Experimental cluster | TCP election + SQL/DDL via log; single-node is **Production GA**. See `docs/en/known-limitations.md`. |
+| Raft consensus | ✅ Supported (3-node, `default` DB) | TCP election + SQL/DDL via log; failover under load, raft TLS, InstallSnapshot recovery e2e-proven. See `docs/en/known-limitations.md`. |
 | Graph / FTS / Columnar | ✅ Implemented | In-memory engines with serialization; FTS/vector/graph indexes persist across restarts. |
 | Query codegen | ✅ Implemented | IR plans compile to storage engine operations with optimization passes. |
 
@@ -1585,10 +1585,10 @@ reflects 100% completion across all major phases.
 
 ## Changelog
 
-See [CHANGELOG.md](CHANGELOG.md) for full release history. Package version is **v1.2.0**.
+See [CHANGELOG.md](CHANGELOG.md) for full release history. Package version is **v1.3.0**.
 
+- **Raft multi-node supported (v1.3.0):** failover under load, mandatory CI gate, raft TLS, InstallSnapshot cold-node recovery — [distributed.md](docs/en/distributed.md), [known-limitations](docs/en/known-limitations.md)
 - **Production GA (single-node):** auth-on prod compose, backup/restore drill, runbook — [known-limitations](docs/en/known-limitations.md), [deployment](docs/en/deployment.md)
-- **Raft multi-node:** experimental — [distributed.md](docs/en/distributed.md)
 
 ## License
 

baradadb.nimble

@@ -1,5 +1,5 @@
 # Package
-version       = "1.2.0"
+version       = "1.3.0"
 author        = "BaraDB Team"
 description   = "BaraDB — Multimodal database written in Nim"
 license       = "BSD-3-Clause"
@@ -30,6 +30,8 @@ task test, "Run all tests":
   for t in ["test_minimal", "test_all", "bugfix_test", "join_tests", "test_lock",
             "test_schema_persist", "test_storage_hardening", "tla_faithfulness",
             "nimforum_smoke_test", "raft_e2e_test", "raft_writes_e2e_test",
+            "raft_failover_load_e2e_test", "raft_tls_e2e_test",
+            "raft_coldnode_e2e_test",
             "fuzz_test", "prop_test",
             "test_wire_insert_stress", "stress_test"]:
     exec "nim c -r tests/" & t & ".nim"

docs/bg/distributed.md

@@ -5,7 +5,7 @@ BaraDB поддържа разпределено внедряване с Raft к
 > ⚠️ **Ограничение при множество бази данни**
 > Разпределените модули (Raft, шардиране и репликация) в момента работят само с **`default`** базата данни. Ако използвате множество бази (`CREATE DATABASE`, `USE DATABASE`), разпределените функции още не ги обхващат. Всяка база данни се нуждае от отделна кластър конфигурация.
 
-> **Статус (2026-07-30):** Raft C3a (мрежова election), C3b (SQL записи), DDL репликация, leader forwarding, log compaction и metrics са **на `main`**. Преглед: `docs/superpowers/specs/2026-07-30-raft-cluster-status.md`.
+> **Статус (2026-07-30, v1.3.0):** Raft е **supported** за обхвата single-`default`-DB: failover под товар, raft TLS и cold-node recovery чрез InstallSnapshot са e2e-доказани. Преглед: `docs/superpowers/specs/2026-07-30-raft-cluster-status.md`.
 
 ## Raft Консенсус
 
@@ -20,10 +20,21 @@ Leader election и log репликация през TCP; SQL DML/DDL за **def
 | `BARADB_RAFT_WRITE_TIMEOUT_MS` | Макс. изчакване за majority commit при SQL записи (по подразбиране 5000) |
 | `BARADB_RAFT_CLIENT_PEERS` | Опционален `id@host:clientPort` map за leader write forwarding |
 | `BARADB_RAFT_LOG_MAX_ENTRIES` | Лимит на in-memory raft log (по подразбиране 256); safe prefix compact |
+| `BARADB_RAFT_SNAP_CHUNK_KB` | Размер на InstallSnapshot chunk в KiB (по подразбиране 256) |
+| `BARADB_RAFT_PEER_STALE_MS` | Peer е „stale“ след толкова ms без ack (по подразбиране 30000); stale peers не блокират compaction |
+| `BARADB_RAFT_TLS_ENABLED` | TLS на raft порта (по подразбиране false; стартът спира при липсващ cert/key) |
+| `BARADB_RAFT_TLS_CERT_FILE` / `BARADB_RAFT_TLS_KEY_FILE` | Сертификат и ключ за raft listener-а |
+| `BARADB_RAFT_TLS_CA_FILE` / `BARADB_RAFT_TLS_VERIFY_PEER` | Опционален CA bundle и mutual auth (по подразбиране false) |
 
 Когато Raft е активен, SQL DML и schema DDL се приемат само от лидера на **`default`**. DML отива като put/delete; DDL — като `ddl` запис. Followers **препращат** write/DDL към лидера, ако е зададен `BARADB_RAFT_CLIENT_PEERS`; иначе връщат `not leader; leader is '…'`. Записи към друга database name се отказват. `CREATE`/`DROP DATABASE` не се репликират. Приложен DML обновява и secondary индекси/графи.
 
-**Log compaction (v1):** след apply node-ът може да изреже safe prefix, когато log-ът надхвърли `BARADB_RAFT_LOG_MAX_ENTRIES`. Leader не реже след matchIndex на peer (catch-up с AppendEntries). Snapshot metadata се пази в `raft_state.bin`.
+**Log compaction:** след apply node-ът може да изреже safe prefix, когато log-ът надхвърли `BARADB_RAFT_LOG_MAX_ENTRIES`. На лидера safe prefix се смята само по peers с ack в рамките на `BARADB_RAFT_PEER_STALE_MS`; stale peers се възстановяват със snapshot при завръщане. Snapshot metadata се пази в `raft_state.bin`.
+
+**Snapshot recovery (InstallSnapshot, v1.3.0):** когато изоставането на follower е невъзстановимо, лидерът изпраща `tar.gz` snapshot на default DB на chunk-ове от `BARADB_RAFT_SNAP_CHUNK_KB` KiB. Follower-ът го възстановява през backup/restore пътя и продължава catch-up. Върнат след дълъг прекъсване или **изтрит** (wiped data dir, същото node id) възел конвергира автоматично. E2E: `tests/raft_coldnode_e2e_test.nim`.
+
+**Клиентски договор при failover:** запис, който е in-flight при смяна на лидера, **гърми бързо с грешка** — клиентът трябва да го повтори (retry). Всеки **потвърден** (acknowledged) запис оцелява failover-а и е наличен на новия лидер и на наваксалите followers. E2E: `tests/raft_failover_load_e2e_test.nim`.
+
+**Raft TLS (v1.3.0):** `BARADB_RAFT_TLS_ENABLED=true` + cert/key на всеки възел; стартът спира при липсващ cert/key. Целият клъстер трябва да е в един и същ режим — plaintext възел не може да говори с TLS порт и е изключен от клъстера (`tests/raft_tls_e2e_test.nim`). Forwarding-ът follower→leader се обвива в TLS, когато клиентският wire порт е с TLS.
 
 **Metrics:** при включен raft `GET /metrics` (HTTP = `BARADB_PORT + 440`) дава Prometheus редове: `baradb_raft_is_leader`, `baradb_raft_term`, `baradb_raft_log_entries`, `baradb_raft_apply_lag`, `baradb_raft_commit_wait_ms_total`, `baradb_raft_elections_total`, `baradb_raft_forwards_total`, `baradb_raft_compactions_total`. `GET /health` включва обект `raft` (`role`, `term`, `leader_id`, …).
 

docs/bg/known-limitations.md

@@ -1,4 +1,4 @@
-# Известни ограничения — v1.2.0 Production GA
+# Известни ограничения — v1.3.0
 
 | Ниво | Значение |
 |------|----------|
@@ -8,25 +8,36 @@
 
 ## Матрица
 
-| Област | GA (v1.2.0) | Experimental / по-късно |
-|--------|-------------|-------------------------|
+| Област | v1.3.0 | Experimental / по-късно |
+|--------|--------|-------------------------|
 | Single-node SQL + LSM | **Supported** | — |
 | Schema / FTS / HNSW / graphs persist | **Supported** | — |
 | Auth + JWT (когато е конфигуриран) | **Supported** | — |
 | Backup / restore | **Supported** | — |
 | Multi-DB (без Raft) | **Supported** | — |
-| Raft 3-node + SQL/DDL | **Experimental** | InstallSnapshot, membership |
+| Raft 3-node (само `default` DB) | **Supported** | failover под товар, TLS, InstallSnapshot — e2e |
 | Raft multi-DB | **Not supported** | само `default` |
+| `CREATE`/`DROP DATABASE` репликация | **Not supported** | per node |
+| Raft membership промени (join/leave) | **Not supported** | фиксиран `BARADB_RAFT_PEERS` |
 | Follower linearizable reads | **Not supported** | best-effort след apply |
+| Rolling upgrades | **Not supported** | рестарт на всички възли заедно — смесени v1.2/v1.3 binaries не трябва да работят в един клъстер |
 | ORC multi-thread shared LSM | **Not supported** | ARC по подразбиране |
 
 ## GA (single-node)
 
 Crash recovery с WAL, schema/index persist, `/health` + `/metrics`, offline backup/restore.
 
-## Raft
+## Raft (supported)
 
-Виж [distributed.md](distributed.md). Staging/ops, **не** v1.2.0 HA продукт.
+Виж [distributed.md](distributed.md). Поддържан обхват: 3-node, DML/DDL само върху `default`, failover под товар (acked writes оцеляват; in-flight writes → грешка, retry), raft TLS, cold-node recovery чрез InstallSnapshot.
+
+## Нови ограничения
+
+- **Legacy REP replication (без raft)** — пътят още извежда delete от празна стойност; insert в PK-only таблица се прилага грешно по него (редът изчезва). Използвай raft.
+- **Snapshot-restore ctx** — след InstallSnapshot restore HTTP endpoints със startup-captured ctx може да сервират стари данни до рестарт (`/query` е свеж per-request); съществуващите клиентски връзки виждат pre-restore състояние — reconnect след restore.
+- **FK-cascade дивергенция под raft** — ефектите на `ON DELETE/UPDATE CASCADE` (и `SET NULL`) не се реплицират през raft: followers прилагат само KV промяната на родителския ред, така че каскадираните дъщерни редове остават на followers. Избягвай FK actions върху raft-реплицирани таблици или приеми периодичен snapshot resync.
+- **Непотвърдени записи в snapshots** — leader прилага записите локално преди raft majority commit; snapshot, направен в този прозорец, може да включи записи, които никога не се commit-ват (фантомни редове след restore + смяна на leadership). Тесен прозорец; поправката е планирана за следващ release.
+- **Блокиране на event loop при snapshot build/restore** — snapshot build/restore изпълнява блокиращ tar/gzip на event loop на възела; големи data dirs могат да забавят heartbeats и да предизвикат election по средата на трансфер.
 
 ## Виж също
 

docs/en/distributed.md

@@ -5,7 +5,7 @@ BaraDB supports distributed deployment with Raft consensus, sharding, and replic
 > ⚠️ **Multi-Database Limitation**
 > The distributed modules (Raft, sharding, and replication) are currently wired to the **`default`** database only. If you use multiple databases (`CREATE DATABASE`, `USE DATABASE`), distributed features do not yet span across them. Each database would need its own cluster setup.
 
-> **Status (2026-07-30):** Raft C3a/C3b + DDL/forward/compact/metrics are **shipped**. Multi-node Raft is **experimental** for v1.2.0 GA (single-node is the production tier). See [known-limitations](known-limitations.md) and `docs/superpowers/specs/2026-07-30-raft-cluster-status.md`.
+> **Status (2026-07-30, v1.3.0):** Raft C3a/C3b + DDL/forward/compact/metrics are **shipped**, and multi-node Raft is **supported** for the single-`default`-DB scope (failover under load, raft TLS, InstallSnapshot cold-node recovery — all e2e-proven). See [known-limitations](known-limitations.md) and `docs/superpowers/specs/2026-07-30-raft-cluster-status.md`.
 
 ## Raft Consensus
 
@@ -20,10 +20,23 @@ Leader election and log replication over TCP; SQL DML/DDL on the default DB go t
 | `BARADB_RAFT_WRITE_TIMEOUT_MS` | Max wait for majority commit on SQL writes (default 5000) |
 | `BARADB_RAFT_CLIENT_PEERS` | Optional `id@host:clientPort` map for leader write forwarding |
 | `BARADB_RAFT_LOG_MAX_ENTRIES` | Soft cap on in-memory raft log length (default 256); safe prefix compact |
+| `BARADB_RAFT_SNAP_CHUNK_KB` | InstallSnapshot chunk size in KiB (default 256) |
+| `BARADB_RAFT_PEER_STALE_MS` | Peer is stale after this many ms without an ack (default 30000); stale peers no longer pin log compaction |
+| `BARADB_RAFT_TLS_ENABLED` | TLS on the raft TCP port (default false; fail-closed startup if cert/key missing) |
+| `BARADB_RAFT_TLS_CERT_FILE` | Server certificate for the raft listener |
+| `BARADB_RAFT_TLS_KEY_FILE` | Private key for the raft listener |
+| `BARADB_RAFT_TLS_CA_FILE` | Optional CA bundle for peer verification |
+| `BARADB_RAFT_TLS_VERIFY_PEER` | Mutual auth — verify client certificates (default false) |
 
 When Raft is enabled, SQL DML (`INSERT`/`UPDATE`/`DELETE`/`MERGE` and transactional `COMMIT`) and schema DDL (`CREATE`/`DROP`/`ALTER` table, index, view, graph, …) are accepted only on the leader of the **`default`** database. DML ships as put/delete log entries; DDL ships as a `ddl` entry with the original SQL and is re-executed on every node at apply. Followers that receive a write/DDL **forward** it to the leader when `BARADB_RAFT_CLIENT_PEERS` maps the leader id to a SQL client address; otherwise they return `not leader; leader is '…'`. Writes against any other database name are rejected (`raft writes only supported on the 'default' database`). `CREATE`/`DROP DATABASE` are not raft-replicated (multi-DB is out of scope for v1). Committed DML also updates secondary B-tree/FTS/HNSW indexes and in-memory graphs.
 
-**Log compaction (v1):** after apply, each node may drop a fully-safe log prefix once `log.len` exceeds `BARADB_RAFT_LOG_MAX_ENTRIES`. The leader never discards past any peer's `matchIndex` (so lagging followers still catch up via AppendEntries). Snapshot metadata (`lastSnapshotIndex`/`Term`) is persisted in `raft_state.bin`; full InstallSnapshot state-machine payloads are not required while this safe-prefix policy holds.
+**Log compaction:** after apply, each node may drop a fully-safe log prefix once `log.len` exceeds `BARADB_RAFT_LOG_MAX_ENTRIES`. On the leader, the safe prefix is computed only over peers that acked within `BARADB_RAFT_PEER_STALE_MS` — stale peers no longer pin compaction and are recovered by snapshot on return. Compaction never goes past `lastApplied`. Snapshot metadata (`lastSnapshotIndex`/`Term`) is persisted in `raft_state.bin`.
+
+**Snapshot recovery (InstallSnapshot, v1.3.0):** when a follower's lag is unrecoverable (the entries it needs were compacted away), the leader builds a `tar.gz` snapshot of the default DB and streams it as `BARADB_RAFT_SNAP_CHUNK_KB`-sized chunks. The follower restores it via the backup/restore path, adopts the snapshot base as its `commitIndex`/`lastApplied`, and resumes normal AppendEntries catch-up. A node that returns after a long outage — and a **wiped** node (data dir deleted, same node id) — both converge automatically through this path. Proven by `tests/raft_coldnode_e2e_test.nim`.
+
+**Client failover contract:** a write that is in flight when the leader dies **fails fast with an error** — the client must retry it (against the new leader, or any follower if `BARADB_RAFT_CLIENT_PEERS` forwarding is configured). Every write the server **acknowledged** survives the failover and is present on the new leader and all caught-up followers. Proven by `tests/raft_failover_load_e2e_test.nim` (leader killed under sustained INSERT load; all acked writes found on both survivors).
+
+**Raft TLS (v1.3.0):** set `BARADB_RAFT_TLS_ENABLED=true` plus `BARADB_RAFT_TLS_CERT_FILE`/`BARADB_RAFT_TLS_KEY_FILE` on every node; startup fails closed if the cert or key is missing. Add `BARADB_RAFT_TLS_CA_FILE` and `BARADB_RAFT_TLS_VERIFY_PEER=true` for mutual authentication. The whole cluster must run the same mode: a plaintext node cannot speak to a TLS port (its frames are undecryptable) and is excluded from the cluster — proven by `tests/raft_tls_e2e_test.nim`. Follower→leader SQL forwarding is TLS-wrapped automatically when the server's client wire port has TLS enabled.
 
 **Metrics:** with raft enabled, `GET /metrics` (HTTP port = `BARADB_PORT + 440`) includes Prometheus lines such as `baradb_raft_is_leader`, `baradb_raft_term`, `baradb_raft_log_entries`, `baradb_raft_apply_lag`, `baradb_raft_commit_wait_ms_total`, `baradb_raft_elections_total`, `baradb_raft_forwards_total`, and `baradb_raft_compactions_total`. `GET /health` embeds a `raft` object (`role`, `term`, `leader_id`, `commit_index`, `apply_lag`, …).
 
@@ -67,6 +80,9 @@ let entry = n1.appendLog("SET key1 value1")
 |------|----------------|
 | `tests/raft_e2e_test.nim` | 3 real processes; election + kill-leader failover |
 | `tests/raft_writes_e2e_test.nim` | DDL/DML via raft, follower forward, index SELECT, failover writes |
+| `tests/raft_failover_load_e2e_test.nim` | Leader killed under sustained write load; every acked write survives |
+| `tests/raft_tls_e2e_test.nim` | Full-TLS 3-node cluster works; plaintext node excluded |
+| `tests/raft_coldnode_e2e_test.nim` | Returning node and wiped node converge via InstallSnapshot |
 
 ## Sharding
 

docs/en/known-limitations.md

@@ -1,6 +1,6 @@
-# Known Limitations — v1.2.0 Production GA
+# Known Limitations — v1.3.0
 
-This page defines **what BaraDB promises** in the v1.2.0 production cut.
+This page defines **what BaraDB promises** in the v1.3.0 production cut.
 
 | Tier | Meaning |
 |------|---------|
@@ -10,20 +10,22 @@ This page defines **what BaraDB promises** in the v1.2.0 production cut.
 
 ## Support matrix
 
-| Area | GA (v1.2.0) | Experimental / later |
-|------|-------------|----------------------|
+| Area | v1.3.0 | Notes |
+|------|--------|-------|
 | Single-node SQL + LSM storage | **Supported** | — |
 | Schema persistence (tables, indexes) | **Supported** | — |
 | FTS / HNSW / graphs across restart | **Supported** | — |
 | Auth + JWT (when configured) | **Supported** | — |
 | Backup / restore (offline, all-databases) | **Supported** | — |
 | Multi-database (non-Raft) | **Supported** | — |
-| Raft 3-node election + SQL/DDL | **Experimental** | InstallSnapshot SM payload, membership |
+| Raft 3-node (single `default` DB) | **Supported** | failover under load, raft TLS, InstallSnapshot recovery — e2e-proven |
+| Leader write forwarding | **Supported** | needs `BARADB_RAFT_CLIENT_PEERS` |
 | Raft multi-database | **Not supported** | only `default` |
-| Leader write forwarding | **Experimental** | needs `BARADB_RAFT_CLIENT_PEERS` |
+| `CREATE`/`DROP DATABASE` replication | **Not supported** | run per node |
+| Raft membership changes (join/leave) | **Not supported** | fixed `BARADB_RAFT_PEERS` set |
 | Follower linearizable reads | **Not supported** | best-effort after apply |
+| Rolling upgrades | **Not supported** | restart all nodes together — mixed v1.2/v1.3 binaries must not run in one cluster |
 | ORC multi-threaded shared LSM | **Not supported** | default is ARC (`nim.cfg`) |
-| Zero-downtime rolling upgrade | **Not supported** | stop → backup → upgrade |
 | Postgres wire protocol | **Not supported** | Bara wire + HTTP |
 
 ## Single-node GA (what you can rely on)
@@ -33,13 +35,22 @@ This page defines **what BaraDB promises** in the v1.2.0 production cut.
 - HTTP `/health` and `/metrics` for process liveness
 - Offline backup of `data/databases` and restore onto an empty data root
 
-## Raft (experimental ops)
+## Raft (supported, single-default-DB scope)
 
-Documented in [distributed.md](distributed.md). Suitable for learning and careful staging; **not** the v1.2.0 HA product tier.
+Documented in [distributed.md](distributed.md). Supported scope:
 
-- SQL DML/DDL on **`default` only**
-- Safe log prefix compact (not full InstallSnapshot)
-- Failover proven in process e2e tests
+- 3-node cluster, SQL DML/DDL on **`default` only**
+- Failover under write load: every acknowledged write survives a leader kill; in-flight writes fail fast — clients must retry
+- TLS on the raft port and on follower→leader forwarding (`BARADB_RAFT_TLS_*`)
+- Cold-node recovery via InstallSnapshot (`BARADB_RAFT_SNAP_CHUNK_KB`, `BARADB_RAFT_PEER_STALE_MS`)
+
+## Newly documented limitations
+
+- **Legacy non-raft REP replication delete inference** — resolved: the non-raft REP payload now carries an explicit put/delete op tag (`encodeRepPayload`/`decodeRepPayload` in `core/replication.nim`), so PK-only inserts (empty LSM value) replicate as puts instead of being misapplied as deletes.
+- **Snapshot-restore ctx staleness** — after an InstallSnapshot restore, HTTP endpoints using the startup-captured ctx may serve stale data until the node is restarted; the `/query` path is fresh per-request. Pre-existing client connections likewise see pre-restore state — reconnect after a restore.
+- **FK-cascade divergence under raft** — `ON DELETE/UPDATE CASCADE` (and `SET NULL`) effects are not raft-replicated: followers only apply the parent row's KV change, so cascaded child rows persist on followers. Avoid FK actions on raft-replicated tables, or accept periodic snapshot resync.
+- **Uncommitted writes in snapshots** — the leader applies writes locally before raft majority commit; a snapshot taken in that window can include writes that never commit (phantom rows after restore + leadership change). Narrow window; fix tracked for a later release.
+- **Event-loop stall during snapshot build/restore** — partially mitigated: the leader's snapshot *send* now tars under the storage gate but runs the CPU-heavy gzip off the event loop on a worker thread (`gzipFileAsync` in `core/backup.nim`), so heartbeats keep firing during compression. The tar itself (send path) and the whole *restore* path (tar extract + DB reopen) still run on the event loop, so very large data dirs can still stall heartbeats during those phases; a full fix (an async/try-lock storage gate so the loop never blocks) is tracked for a later release.
 
 ## Operational requirements
 

docs/en/release-checklist.md

@@ -1,4 +1,4 @@
-# Release checklist — v1.2.0 Production GA
+# Release checklist — v1.3.0 raft-supported
 
 Use before tagging and publishing artifacts.
 
@@ -6,8 +6,8 @@ Use before tagging and publishing artifacts.
 
 - [ ] Working tree clean on `main`
 - [ ] [Known limitations](known-limitations.md) accurate
-- [ ] `CHANGELOG.md` has dated `## [1.2.0]` (not Unreleased for shipped items)
-- [ ] `baradadb.nimble` version `1.2.0`
+- [ ] `CHANGELOG.md` has dated `## [1.3.0]` (not Unreleased for shipped items)
+- [ ] `baradadb.nimble` version `1.3.0`
 
 ## Tests
 
@@ -23,9 +23,12 @@ nim c -d:ssl --threads:on --path:src -r tests/test_schema_persist.nim
 ./scripts/backup-restore-drill.sh
 DRILL_PORT=19482 ./scripts/backup-restore-drill.sh
 
-# Optional cluster e2e (experimental tier)
+# Cluster e2e (raft supported tier — all five suites)
 ./tests/raft_e2e_test
 ./tests/raft_writes_e2e_test
+./tests/raft_failover_load_e2e_test
+./tests/raft_tls_e2e_test
+./tests/raft_coldnode_e2e_test
 ```
 
 ## Production compose
@@ -41,17 +44,17 @@ docker compose -f docker-compose.prod.yml config >/dev/null
 
 ```bash
 nimble build_release   # or: nim c -d:release -o:build/baradadb src/baradadb.nim
-docker build -t baradb:1.2.0 -t baradb:latest .
+docker build -t baradb:1.3.0 -t baradb:latest .
 ```
 
 ## Tag
 
 ```bash
-git tag -a v1.2.0 -m "BaraDB v1.2.0 Production GA (single-node)"
+git tag -a v1.3.0 -m "BaraDB v1.3.0 raft-supported"
 git push origin main --tags
 ```
 
 ## Post-release
 
 - [ ] Smoke: start prod compose, `/health` → ok, auth required for `/query`
-- [ ] Announce: single-node GA; Raft experimental (link known-limitations)
+- [ ] Announce: raft-supported release (3-node, `default` DB); link known-limitations

docs/superpowers/plans/2026-07-30-v1.3.0-raft-supported.md

@@ -0,0 +1,598 @@
+# v1.3.0 Raft-Supported — Implementation Plan
+
+> **For agentic workers:** REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (`- [ ]`) syntax for tracking.
+> Spec first: `docs/superpowers/specs/2026-07-30-raft-supported-design.md`.
+
+**Goal:** Move raft from experimental to supported: proven failover under
+load, mandatory CI e2e, cold-node recovery via InstallSnapshot, raft-port TLS.
+
+**Architecture:** No changes to election/AppendEntries semantics. New
+InstallSnapshot message pair rides the existing framed TCP transport
+(backward-compatible trailing fields, `RaftProtoVersion` stays 1). TLS wraps
+the existing transport via `protocol/ssl.nim`. Snapshot payload reuses
+`core/backup.nim` tar.gz backup/restore.
+
+**Tech stack:** Nim 2.2.x, std/asyncnet + std/net SSL, existing
+`tests/raft_*_e2e_test.nim` process harness, GitHub Actions.
+
+## Global Constraints
+
+- Spec: `docs/superpowers/specs/2026-07-30-raft-supported-design.md`.
+- Do **not** change election safety, commit rules, or the DDL/DML
+  classification from C3b/C3c.
+- Raft remains default-DB-only; snapshot payload covers the default DB only.
+- Compile all touched Nim with `-d:ssl --threads:on --path:src`.
+- Test baseline per task: `tests/test_all.nim` + `tests/bugfix_test.nim` must
+  stay green; raft e2e suites green where the binary is built.
+- Branch: `main` (short-lived feature branches merged same day are fine).
+- Version bump to 1.3.0 happens only in the final task (T12).
+
+---
+
+## Phase map
+
+| Phase | Tasks | Outcome |
+|-------|-------|---------|
+| P1 Proof | T1 | Failover-under-load e2e |
+| P2 CI | T2 | Mandatory raft e2e CI gate |
+| P3 TLS | T3–T6 | Raft port TLS + mutual auth + TLS e2e |
+| P4 Cold node | T7–T11 | InstallSnapshot + compaction unpin + cold-node e2e |
+| P5 Release | T12 | Docs, limitations, version 1.3.0 |
+
+---
+
+### Task 1: Failover-under-load E2E
+
+**Files:**
+- Create: `tests/raft_failover_load_e2e_test.nim`
+- Modify: `baradadb.nimble` (task `test`, line ~30-34: add `raft_failover_load_e2e_test` after `raft_writes_e2e_test`)
+
+**Interfaces:**
+- Consumes: process harness conventions from `tests/raft_writes_e2e_test.nim`
+  (`NodeProc`, `drainOutput`, `portOpen`, `openClient`, `waitForRow` pattern);
+  client `adaptors/nim/baradb_sqlite`.
+- Produces: suite `Raft failover under load E2E`.
+
+**Scenario (spec D1):**
+- Port base `cbase = 50000 + (tstamp mod 4000)`, `rbase = cbase + 100`
+  (distinct from 35000/41000/46000 bases already in use).
+- Boot 3 nodes with `BARADB_RAFT_*` env exactly as
+  `raft_writes_e2e_test.nim:150-171`.
+- Wait for stable leader (same `maxLeader` logic). Leader DDL:
+  `CREATE TABLE load_test (id INT PRIMARY KEY)`.
+- Load phase: spawn a Nim `Thread` that loops `n = 1, 2, ...`:
+  `INSERT INTO load_test (id) VALUES (n)` against the current leader's
+  client port; every **acknowledged** `n` appended to a
+  `seq[int]` guarded by a `Lock`. On exception: reopen client against a
+  survivor, continue (this models the documented client retry contract).
+- At ≥ 50 acked writes: `killNode(leader)`.
+- Assert A (availability): some survivor accepts an INSERT within 10 s of
+  the kill.
+- Assert B (durability): after new leader is stable and the remaining
+  follower caught up (poll `SELECT count(*)` equality or 10 s deadline),
+  `SELECT id FROM load_test` on **both** survivors contains every acked id.
+- Stop the writer thread in `finally`; reuse the `dumpAll`-on-fail
+  convention.
+
+- [x] **Step 1:** Write the suite skeleton: harness copied from
+  `raft_writes_e2e_test.nim` (drain/kill/leader-discovery helpers), writer
+  thread, kill at 50 acked, asserts A + B.
+- [x] **Step 2:** Build binary and run:
+  `nim c -o:build/baradadb src/baradadb.nim && nim c -d:ssl --threads:on --path:src -r tests/raft_failover_load_e2e_test.nim`
+  Expected: PASS.
+- [x] **Step 3:** Run 3 consecutive times (failover timing flakiness check).
+- [x] **Step 4:** Add suite to `nimble test` list in `baradadb.nimble`.
+- [x] **Step 5:** Commit
+  `test(raft): failover under sustained write load e2e`
+
+---
+
+### Task 1a: Fix raft put/delete encoding for empty values (bug found in T1)
+
+**Bug:** `execInsert` (`src/barabadb/query/exec/dml.nim:60-90`) stores only
+non-PK columns in the value, so a PK-only table row gets `valStr = ""` and
+`kvPairs.add((fullKey, @[]))`. `appendWriteToRaft`
+(`src/barabadb/core/server.nim:309-330`) encodes an empty value as a
+`"delete"` log entry — but `execDelete` (`dml.nim:223-241`) uses the same
+`(fullKey, @[])` shape for real deletes. Result: INSERT into a PK-only
+table returns OK after majority commit, then every node (leader included,
+on apply) **deletes the row**. Verified live in T1: 30 acked inserts → 0
+rows on all nodes.
+
+**Files:**
+- Modify: `src/barabadb/query/exec/types.nim:139` — `ExecResult.keyValuePairs`
+- Modify: `src/barabadb/query/exec/dml.nim` — 3 producer sites (insert ~90,
+  delete ~241, update ~316)
+- Modify: `src/barabadb/core/server.nim` — `appendWriteToRaft` (~309) and
+  its call site (~441-464)
+- Test: `tests/bugfix_test.nim` (new suite)
+
+**Interfaces:**
+- Change the pair type to carry the op explicitly:
+
+```nim
+# exec/types.nim
+keyValuePairs*: seq[tuple[key: string, value: seq[byte], deleted: bool]]
+```
+
+- `execInsert`/`execUpdate` produce `deleted: false` (even when
+  `value.len == 0`); `execDelete` produces `deleted: true`.
+- `appendWriteToRaft` encodes `deleted` → `"delete"`, else `"put"`
+  (empty value stays a put). Apply side (`baradadb.nim:358-368`) already
+  handles `put` with empty value correctly — no change needed there.
+- Check other `keyValuePairs` consumers compile clean (replication path in
+  `server.nim`); keep `okResult(kvPairs=...)` call sites type-correct.
+
+**Steps:**
+- [x] **Step 1:** Write the failing test in `tests/bugfix_test.nim`:
+  build `ExecResult` via the insert path for a PK-only table (or call
+  `appendWriteToRaft` semantics directly): assert a PK-only insert yields
+  a pair with `deleted == false` and encodes as `"put"`, while a delete
+  yields `deleted == true` and encodes as `"delete"`.
+- [x] **Step 2:** Run, expect fail/compile error.
+- [x] **Step 3:** Implement the type + producer/consumer changes.
+- [x] **Step 4:** `bugfix_test` + `test_all` green; rebuild
+  `build/baradadb` and re-run `tests/raft_failover_load_e2e_test.nim` —
+  then **switch its table back** to the brief's original
+  `load_test (id INT PRIMARY KEY)` / `VALUES (n)` shape (remove the
+  two-column workaround and its header note) and re-run green.
+- [x] **Step 5:** Commit
+  `fix(raft): distinguish put-with-empty-value from delete in write path`
+
+---
+
+### Task 2: Mandatory raft e2e CI gate
+
+**Files:**
+- Modify: `.github/workflows/ci.yml`
+- Modify: `tests/raft_e2e_test.nim`, `tests/raft_writes_e2e_test.nim`,
+  `tests/raft_failover_load_e2e_test.nim` (skip→fail under CI)
+
+**Steps:**
+- [x] **Step 1:** In each suite's binary-missing branch, replace plain
+  `skip()` with:
+
+```nim
+if not fileExists(BinaryPath):
+  if getEnv("CI").len > 0:
+    echo "[FAIL] ", BinaryPath, " missing under CI — build step broken?"
+    fail()
+  else:
+    echo "[SKIP] ", BinaryPath, " missing — run `nimble test` (builds the server first)"
+    skip()
+```
+
+- [x] **Step 2:** Add a dedicated job to `.github/workflows/ci.yml` (after
+  the `test` job), modeled on its setup steps:
+
+```yaml
+  raft-e2e:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - name: Setup Nim
+        uses: jiro4989/setup-nim-action@v1
+        with:
+          nim-version: '2.2.10'
+      - name: Install system dependencies
+        run: sudo apt-get update -qq && sudo apt-get install -y -qq libssl-dev libpcre3-dev openssl ca-certificates
+      - name: Install Nim dependencies
+        run: nimble install --depsOnly -y
+      - name: Build server
+        run: nim c -d:ssl -o:build/baradadb src/baradadb.nim
+      - name: Raft e2e suites
+        env:
+          CI: "true"
+        run: |
+          nim c -d:ssl --threads:on --path:src -r tests/raft_e2e_test.nim
+          nim c -d:ssl --threads:on --path:src -r tests/raft_writes_e2e_test.nim
+          nim c -d:ssl --threads:on --path:src -r tests/raft_failover_load_e2e_test.nim
+```
+
+- [x] **Step 3:** Push on a branch; confirm the `raft-e2e` job appears and
+  is green; confirm deleting `build/` would fail (local run with `CI=true`
+  and no binary → FAIL).
+- [x] **Step 4:** Commit
+  `ci(raft): dedicated mandatory raft e2e job; no silent skips under CI`
+
+---
+
+### Task 3: Raft TLS config + fail-closed startup
+
+**Files:**
+- Modify: `src/barabadb/core/config.nim` (fields after `raftLogMaxEntries`
+  at lines ~46/88; env parsing after line ~212)
+- Modify: `src/baradadb.nim` (raft wiring block, lines 337-377)
+- Test: `tests/bugfix_test.nim` (new suite)
+
+**Interfaces:**
+- Produces config fields (used by T4/T6):
+
+```nim
+raftTlsEnabled*: bool       # default false
+raftTlsCertFile*: string    # default ""
+raftTlsKeyFile*: string     # default ""
+raftTlsCaFile*: string      # default ""
+raftTlsVerifyPeer*: bool    # default false
+```
+
+- Env parsing (mirror lines 184-212 style):
+
+```nim
+cfg.raftTlsEnabled = parseEnvBool(getEnv("BARADB_RAFT_TLS_ENABLED", ""), cfg.raftTlsEnabled)
+cfg.raftTlsCertFile = getEnv("BARADB_RAFT_TLS_CERT_FILE", cfg.raftTlsCertFile)
+cfg.raftTlsKeyFile = getEnv("BARADB_RAFT_TLS_KEY_FILE", cfg.raftTlsKeyFile)
+cfg.raftTlsCaFile = getEnv("BARADB_RAFT_TLS_CA_FILE", cfg.raftTlsCaFile)
+cfg.raftTlsVerifyPeer = parseEnvBool(getEnv("BARADB_RAFT_TLS_VERIFY_PEER", ""), cfg.raftTlsVerifyPeer)
+```
+
+- Fail-closed in `baradadb.nim` before `newRaftNetwork`:
+
+```nim
+if config.raftTlsEnabled:
+  if config.raftTlsCertFile.len == 0 or config.raftTlsKeyFile.len == 0 or
+     not fileExists(config.raftTlsCertFile) or not fileExists(config.raftTlsKeyFile):
+    raise newException(ValueError,
+      "BARADB_RAFT_TLS_ENABLED=true but cert/key missing " &
+      "(BARADB_RAFT_TLS_CERT_FILE / BARADB_RAFT_TLS_KEY_FILE)")
+```
+
+- [x] **Step 1:** Write failing test in `tests/bugfix_test.nim`: default
+  config has `raftTlsEnabled == false`; env `BARADB_RAFT_TLS_ENABLED=true`
+  + cert paths parse into config.
+- [x] **Step 2:** Run test, expect compile/fail (fields don't exist).
+- [x] **Step 3:** Implement fields + env parsing + startup check.
+- [x] **Step 4:** Test green; `test_all` still green.
+- [x] **Step 5:** Commit
+  `feat(raft): TLS config surface with fail-closed startup`
+
+---
+
+### Task 4: TLS in RaftNetwork transport
+
+**Files:**
+- Modify: `src/barabadb/core/raft.nim` (`RaftNetwork` type ~line 731,
+  `connectToPeer` ~748, `run` ~857)
+- Modify: `src/baradadb.nim` (construct TLSContext, pass to network)
+- Test: `tests/test_all.nim` (in-process TLS raft pair)
+
+**Interfaces:**
+- Consumes: `protocol/ssl.nim` — `newTLSConfig(certFile, keyFile, caFile,
+  verifyPeer)`, `newTLSContext`, `wrapClient`, `wrapServer`.
+- Produces: `RaftNetwork.tls*: TLSContext` (nil = plaintext, unchanged
+  default); `newRaftNetwork(node, tls = nil)`.
+
+**Implementation:**
+
+```nim
+# raft.nim — in connectToPeer, after successful connect:
+if net.tls != nil:
+  try: net.tls.wrapClient(sock)
+  except CatchableError:
+    try: sock.close() except CatchableError: discard
+    return
+
+# raft.nim — in run, after accept, before receiveLoop:
+if net.tls != nil:
+  try: net.tls.wrapServer(client)
+  except CatchableError:
+    client.close()
+    continue
+```
+
+- `baradadb.nim`: build the context when enabled:
+
+```nim
+var raftTls: TLSContext = nil
+if config.raftTlsEnabled:
+  raftTls = newTLSContext(newTLSConfig(
+    config.raftTlsCertFile, config.raftTlsKeyFile,
+    caFile = config.raftTlsCaFile, verifyPeer = config.raftTlsVerifyPeer))
+...
+raftNet = newRaftNetwork(raftNode, raftTls)
+```
+
+- [x] **Step 1:** Write failing in-process test (`test_all.nim`): two
+  `RaftNode`s over `RaftNetwork` with a self-signed cert from
+  `generateSelfSignedCert` (`protocol/ssl.nim:79`) — election completes
+  over TLS; plaintext dial to the TLS port produces no protocol effect
+  (no state change, connection dropped).
+- [x] **Step 2:** Run, expect fail (no `tls` field).
+- [x] **Step 3:** Implement transport changes + wiring.
+- [x] **Step 4:** Test green; plaintext raft e2e suites still green
+  (regression: nil-TLS path untouched).
+- [x] **Step 5:** Commit
+  `feat(raft): optional TLS on raft transport (server + dialer)`
+
+---
+
+### Task 5: TLS for leader SQL forwarding
+
+**Files:**
+- Modify: `src/barabadb/core/server.nim` (`forwardQueryToLeader`, lines
+  210-289)
+
+**Interfaces:**
+- Consumes: `protocol/ssl.nim` `wrapClient`; server config `tlsEnabled`,
+  `certFile`, `keyFile`.
+- Produces: `forwardQueryToLeader(host, port, query, tls: TLSContext = nil,
+  ...)`.
+
+- [x] **Step 1:** When the server's client wire port has TLS on
+  (`server.tls != nil`), wrap the forwarding socket with a client-side
+  context before sending the wire header; on handshake failure return
+  `(false, QueryResult(), "leader forward TLS handshake failed")`.
+- [x] **Step 2:** Manual check: TLS server + raft forwarding (follower
+  INSERT forwarded over TLS) works; non-TLS setup unchanged.
+- [x] **Step 3:** Commit
+  `feat(raft): TLS on follower→leader SQL forwarding`
+
+---
+
+### Task 6: Raft TLS E2E
+
+**Files:**
+- Create: `tests/raft_tls_e2e_test.nim`
+- Modify: `baradadb.nimble` (test list), `.github/workflows/ci.yml`
+  (raft-e2e job: add this suite)
+
+**Interfaces:**
+- Consumes: T3/T4 implementation; harness from T1; openssl CLI for cert
+  generation (already used by `protocol/ssl.nim`).
+
+**Scenario:**
+- Port base `54000 + (tstamp mod 4000)`.
+- Generate one self-signed cert per node into the temp data dirs
+  (`openssl req -x509 ...`, or `generateSelfSignedCert`).
+- Boot 3 nodes with `BARADB_RAFT_TLS_ENABLED=true` + per-node cert/key;
+  assert election + `CREATE TABLE` via raft DDL + one replicated INSERT
+  visible on a follower.
+- Negative: start a 4th process with raft TLS **disabled** pointed at the
+  same peers; assert the TLS cluster still elects/operates among its 3
+  members and the plaintext node never becomes leader (its frames are
+  undecryptable).
+- Same `CI`-fail semantics as T2.
+
+- [x] **Step 1:** Write suite; Step 2: run green locally; Step 3: run 3×
+  (timing); Step 4: wire into `nimble test` + CI job; Step 5: Commit
+  `test(raft): 3-node TLS cluster e2e with plaintext rejection`
+
+---
+
+### Task 7: InstallSnapshot protocol
+
+**Files:**
+- Modify: `src/barabadb/core/raft.nim` (`RaftMessageKind` ~line 80,
+  `RaftMessage` ~86, `serialize` ~679, `deserializeRaftMessage` ~702)
+- Test: `tests/test_all.nim` (serialize/deserialize round-trip)
+
+**Interfaces:**
+- Produces:
+
+```nim
+# RaftMessageKind += rmkInstallSnapshot, rmkInstallSnapshotReply
+# RaftMessage new fields:
+snapId*: uint64      # snapshot generation, matches leader's base at build time
+snapOffset*: uint64  # byte offset of this chunk within the archive
+snapData*: seq[byte] # chunk payload (<= snapChunkBytes)
+snapDone*: bool      # last chunk
+# Reused for this kind: prevLogIndex = snapshot base index,
+# prevLogTerm = snapshot base term. Reply uses success/matchIdx as usual.
+```
+
+- Serialization: append `snapId`, `snapOffset`, `snapData` (length-prefixed),
+  `snapDone` **after** `matchIdx`; deserialize each with `if not s.atEnd`
+  guards (pattern from `loadState`, raft.nim:163-167). Old binaries ignore
+  trailing bytes; new binaries default missing fields to zero/false.
+  `RaftProtoVersion` stays 1.
+
+- [x] **Step 1:** Write failing round-trip test: all new fields survive
+  serialize→deserialize; a buffer serialized by the *old* layout (no
+  trailing fields) deserializes with zero defaults.
+- [x] **Step 2:** Run, expect fail.
+- [x] **Step 3:** Implement.
+- [x] **Step 4:** Green; existing raft suites still green.
+- [x] **Step 5:** Commit
+  `feat(raft): InstallSnapshot wire protocol (backward-compatible)`
+
+---
+
+### Task 8: Follower snapshot receive + restore
+
+**Files:**
+- Modify: `src/barabadb/core/raft.nim` (`RaftNode` — new callback +
+  incoming-snapshot buffer; `processMessage` ~786)
+- Modify: `src/barabadb/core/config.nim` (`raftSnapChunkKb: int`, env
+  `BARADB_RAFT_SNAP_CHUNK_KB`, default 256; parsing next to the other
+  `BARADB_RAFT_*` env reads)
+- Modify: `src/baradadb.nim` (wire `restoreSnapshot` callback using
+  `core/backup.nim` + `DatabaseRegistry`; pass chunk size to the node)
+- Test: `tests/test_all.nim`
+
+**Interfaces:**
+- Produces on `RaftNode`:
+
+```nim
+snapChunkBytes*: int          # from BARADB_RAFT_SNAP_CHUNK_KB, default 262144
+restoreSnapshot*: proc(archivePath: string, baseIndex: uint64,
+                       baseTerm: uint64): bool {.gcsafe.}
+snapIncomingId*: uint64
+snapIncomingFile*: string     # temp path under dataDir/raft/snap_incoming/
+```
+
+- `processMessage` case `rmkInstallSnapshot`: append `snapData` at
+  `snapOffset` to the temp file (create/truncate when `snapId !=
+  snapIncomingId`); on `snapDone`: call `restoreSnapshot`; on success set
+  `lastSnapshotIndex/Term = prevLogIndex/prevLogTerm`,
+  `commitIndex = lastApplied = lastSnapshotIndex`, clear `log`,
+  `saveState()`, reply success with `matchIdx = lastSnapshotIndex`; on
+  failure reply `success = false` and delete the temp file.
+- `baradadb.nim` `restoreSnapshot` implementation: close default DB via
+  registry, `restoreDataDir(archivePath, defaultDbDir)`
+  (`backup.nim:263`), reopen, swap `ctx`. Return false on any exception.
+
+- [x] **Step 1:** Write failing test: feed a node two chunks + done with a
+  real tar.gz fixture; assert callback received the assembled file, state
+  fields updated, log cleared.
+- [x] **Step 2:** Run, expect fail. **Step 3:** Implement.
+- [x] **Step 4:** Green + regression suites. **Step 5:** Commit
+  `feat(raft): follower InstallSnapshot receive and restore`
+
+---
+
+### Task 9: Leader snapshot send
+
+**Files:**
+- Modify: `src/barabadb/core/raft.nim` (`handleAppendReply` floor branch
+  ~526-531, new `sendSnapshot` proc, per-peer reject counter)
+- Modify: `src/baradadb.nim` (wire `buildSnapshot` callback)
+- Test: `tests/test_all.nim`
+
+**Interfaces:**
+- Consumes: T7 protocol, `backupDataDir` (`backup.nim:225`).
+- Produces on `RaftNode`:
+
+```nim
+buildSnapshot*: proc(destPath: string): bool {.gcsafe.}
+snapRejectStreak*: Table[string, int]   # consecutive floor-level rejects per peer
+```
+
+- Logic: in `handleAppendReply`, when a reject arrives **and**
+  `nextIndex[peerId] == lastSnapshotIndex + 1` (floor reached): increment
+  streak; at streak ≥ 2 the leader knows the follower needs a snapshot →
+  `asyncCheck sendSnapshot(peerId)`. Reset streak on any successful reply.
+- `sendSnapshot`: `buildSnapshot` into `dataDir/raft/snap_out_<snapId>.tar.gz`
+  (`snapId = lastSnapshotIndex`); stream chunks of `snapChunkBytes` as
+  `rmkInstallSnapshot`; on final success reply set
+  `matchIndex[peer] = lastSnapshotIndex`,
+  `nextIndex[peer] = lastSnapshotIndex + 1`; delete the temp archive.
+
+- [x] **Step 1:** Write failing test: leader with compacted log
+  (`lastSnapshotIndex = 100`) + peer at floor rejecting twice → snapshot
+  messages emitted; success reply advances `matchIndex`/`nextIndex`.
+- [x] **Step 2:** Run, expect fail. **Step 3:** Implement.
+- [x] **Step 4:** Green + regression. **Step 5:** Commit
+  `feat(raft): leader InstallSnapshot send on unrecoverable lag`
+
+---
+
+### Task 10: Unpin compaction from dead peers
+
+**Files:**
+- Modify: `src/barabadb/core/raft.nim` (`compactLog` ~244, `becomeLeader`
+  ~327, `handleAppendReply` ~487)
+- Modify: `src/barabadb/core/config.nim` (`raftPeerStaleMs`, env
+  `BARADB_RAFT_PEER_STALE_MS`, default 30000)
+- Test: `tests/test_all.nim`
+
+**Interfaces:**
+- Produces: `matchIndexSeenMs*: Table[string, int64]` on `RaftNode` —
+  monotonic ms timestamp of the last successful reply per peer, updated in
+  `handleAppendReply` success branch and initialized to "now" in
+  `becomeLeader`.
+
+**Logic:** leader-side `compactLog` computes `minMatch` only over peers
+with `now - matchIndexSeenMs[peer] <= raftPeerStaleMs`; peers stale longer
+are excluded (they'll be snapshotted on return per T9). Follower path
+unchanged. Guard: never compact past `lastApplied`.
+
+- [x] **Step 1:** Write failing test: leader, one peer never replies,
+  log > maxEntries → with default stale window, log compacts through
+  lastApplied anyway; with the peer responsive, compaction still pins at
+  its matchIndex (existing safety preserved).
+- [x] **Step 2:** Run, expect fail. **Step 3:** Implement.
+- [x] **Step 4:** Green + regression. **Step 5:** Commit
+  `feat(raft): compaction unpinned from stale peers (snapshot fallback)`
+
+---
+
+### Task 11: Cold-node E2E
+
+**Files:**
+- Create: `tests/raft_coldnode_e2e_test.nim`
+- Modify: `baradadb.nimble`, `.github/workflows/ci.yml` (raft-e2e job)
+
+**Interfaces:** Consumes T7–T10; harness from T1. Port base
+`58000 + (tstamp mod 4000)`. Small `BARADB_RAFT_LOG_MAX_ENTRIES=16` and
+`BARADB_RAFT_PEER_STALE_MS=3000` to force compaction quickly.
+
+**Scenario A — node returns after compaction:**
+1. 3-node cluster, create table, kill node n3.
+2. Write 100 rows through the leader (forces compaction past n3's
+   matchIndex once n3 is stale).
+3. Assert via `/metrics` on the leader HTTP port
+   (`baradb_raft_log_entries`) that the log stayed bounded.
+4. Restart n3 with its intact data dir; assert it receives a snapshot
+   (leader log line / `baradb_raft_snapshot_index` advances on n3) and
+   within 15 s `SELECT count(*)` on n3 matches the leader.
+
+**Scenario B — wiped node joins:**
+1. Stop n3, **delete its data dir**, restart with the same node id.
+2. Assert it converges (snapshot → catch-up) and serves the full row set
+   within 20 s.
+
+- [x] **Step 1:** Write suite. **Step 2:** Green locally. **Step 3:** 3×
+  stability runs. **Step 4:** `nimble test` + CI wiring. **Step 5:** Commit
+  `test(raft): cold-node rejoin and wiped-node join e2e`
+
+---
+
+### Task 12: Docs, limitations, version 1.3.0
+
+**Files:**
+- Modify: `docs/en/distributed.md`, `docs/bg/distributed.md` — client
+  failover contract (in-flight writes fail fast, retry; acked writes
+  durable), TLS setup (`BARADB_RAFT_TLS_*`), snapshot behavior/tunables
+  (`BARADB_RAFT_SNAP_CHUNK_KB`, `BARADB_RAFT_PEER_STALE_MS`)
+- Modify: `docs/en/known-limitations.md`, `docs/bg/known-limitations.md` —
+  raft 3-node moves from "Experimental" to "Supported" for the covered
+  scope; remaining non-goals (multi-DB raft, membership changes, read
+  consistency levels) stay listed
+- Modify: `docs/superpowers/specs/2026-07-30-raft-cluster-status.md` —
+  status → v1.3.0 supported
+- Modify: `CHANGELOG.md` — `## [1.3.0] — <ship date>`
+- Modify: `baradadb.nimble` → `version = "1.3.0"`; README status lines
+- Modify: `docs/en/release-checklist.md` — add raft TLS + cold-node suites
+
+- [x] **Step 1:** Doc edits. **Step 2:** Full `nimble test` green.
+- [x] **Step 3:** Commit
+  `release: v1.3.0 raft-supported (failover load, CI gate, snapshot, TLS)`
+- [ ] **Step 4 (human/controller):** tag `v1.3.0` after review.
+
+---
+
+## Task dependency graph
+
+```
+T1 failover-load e2e ──→ T2 CI gate
+T3 TLS config ──→ T4 transport TLS ──→ T5 forward TLS ──→ T6 TLS e2e
+T7 snapshot protocol ──→ T8 follower restore ──→ T9 leader send ──→ T10 unpin ──→ T11 cold-node e2e
+all ──→ T12 docs/version
+```
+
+## Explicit out-of-scope
+
+- Membership change (join/leave) protocol
+- Multi-database raft; `CREATE`/`DROP DATABASE` replication
+- Linearizable follower reads
+- Rolling-upgrade compat shims beyond the trailing-field guard
+  (upgrade = restart all nodes)
+
+## Definition of Done
+
+- [x] All P1–P5 tasks complete, `nimble test` green
+- [ ] `raft-e2e` CI job green and mandatory (no silent skip)
+- [x] Failover-under-load e2e: every acked write survives leader kill
+- [x] Cold-node e2e: returning node and wiped node converge automatically
+- [x] Raft TLS e2e: full-TLS cluster works; plaintext node excluded
+- [x] known-limitations updated: raft supported for the covered scope
+
+## Estimated effort
+
+| Phase | Effort |
+|-------|--------|
+| P1–P2 | 0.5–1 day |
+| P3 | 1 day |
+| P4 | 2–3 days |
+| P5 | 0.5 day |
+| **Total** | **~4–5 focused days** |

docs/superpowers/specs/2026-07-30-raft-cluster-status.md

@@ -1,9 +1,38 @@
-# Raft Cluster Status — C3a / C3b / post-C3b
+# Raft Cluster Status — C3a / C3b / post-C3b / v1.3.0
 
 Date: 2026-07-30  
-Status: **Shipped on `main`** (tip includes metrics).  
+Status: **v1.3.0 — Supported** for the single-`default`-DB scope (see the v1.3.0 section below).  
 Branch: all work merged to `main` only (feature branch removed).
 
+## v1.3.0 — raft-supported (2026-07-30)
+
+Raft moves from Experimental to **Supported** for a 3-node cluster on the
+`default` database. Landed on top of the C3a/C3b base:
+
+- **Failover under load** — `tests/raft_failover_load_e2e_test.nim`: leader
+  killed under sustained writes; every acked write survives (client contract:
+  in-flight writes fail fast, retry).
+- **Mandatory CI gate** — dedicated `raft-e2e` job runs all five raft e2e
+  suites; missing binary is a hard FAIL under CI.
+- **Raft TLS** — `BARADB_RAFT_TLS_*` config, fail-closed startup, optional
+  mutual auth, TLS on follower→leader forwarding;
+  `tests/raft_tls_e2e_test.nim` (plaintext node excluded).
+- **InstallSnapshot** — backward-compatible wire protocol, leader chunk send
+  (`BARADB_RAFT_SNAP_CHUNK_KB`, default 256), follower restore via
+  backup/restore, compaction unpinned from stale peers
+  (`BARADB_RAFT_PEER_STALE_MS`, default 30000);
+  `tests/raft_coldnode_e2e_test.nim` (returning + wiped node converge).
+- **Fixes** — put/delete encoding (`deleted` flag), rejoin livelock (cached
+  peer sockets dropped on leadership), post-restore ctx repoint.
+
+Resolved non-goals from the list below: raft-port TLS, InstallSnapshot with
+full SM payload. Still open: multi-database raft, `CREATE`/`DROP DATABASE`
+replication, membership changes, linearizable follower reads, rolling
+upgrades (restart all nodes together).
+
+Plan: `docs/superpowers/plans/2026-07-30-v1.3.0-raft-supported.md` ·
+Design: `docs/superpowers/specs/2026-07-30-raft-supported-design.md`
+
 ## Phase map
 
 | Phase | Spec / plan | Status | What landed |

docs/superpowers/specs/2026-07-30-raft-supported-design.md

@@ -0,0 +1,212 @@
+# v1.3.0 Raft-Supported — Design Spec
+
+Date: 2026-07-30
+Status: Draft
+Follows: `2026-07-30-raft-cluster-status.md` (C3a/C3b/C3c-lite shipped),
+`2026-07-30-production-ga-design.md` ("After GA" section).
+
+## Goal
+
+Move the raft cluster from **experimental** to **supported** by closing the
+four gaps named in the GA plan: failover under load (proven, not assumed),
+CI e2e mandatory, a cold-node story, and raft-port TLS.
+
+Non-goals (unchanged from C3 status doc): multi-database raft, membership
+change protocol, read consistency levels, `CREATE`/`DROP DATABASE`
+replication.
+
+## Current state (verified 2026-07-30)
+
+- `src/barabadb/core/raft.nim` (919 lines): election, AppendEntries,
+  safe-prefix compaction, metrics, plain-TCP `RaftNetwork` transport.
+- Wiring: `src/baradadb.nim:337-377` (env `BARADB_RAFT_*`, state in
+  `dataDir/raft/raft_state.bin`).
+- Leader forwarding: `src/barabadb/core/server.nim:210-289`
+  (`forwardQueryToLeader`, plain TCP).
+- TLS infra exists for the client wire port only:
+  `src/barabadb/protocol/ssl.nim` (`TLSConfig`, `TLSContext`, `wrapClient`,
+  `wrapServer`); server accept loop wraps at `core/server.nim:876-889`.
+- E2E: `tests/raft_e2e_test.nim` (election + failover),
+  `tests/raft_writes_e2e_test.nim` (DDL/DML replication, forwarding,
+  failover write probe). Both run under `nimble test`, which CI runs.
+
+## Gap analysis
+
+### G1. Failover under load — unproven
+
+The existing failover test (`raft_writes_e2e_test.nim:349-405`) kills the
+leader *while idle* and probes a single INSERT afterwards. Nothing tests a
+write workload running *during* the leader crash, and nothing verifies that
+every client-acknowledged write survives the failover (raft's core promise:
+committed entries are never lost).
+
+### G2. CI e2e — present but silently skippable
+
+Both e2e suites `skip()` when `./build/baradadb` is missing
+(`raft_writes_e2e_test.nim:413-417`). `nimble test` builds the binary first,
+so CI runs them today — but a broken build step or a renamed binary turns a
+raft regression into a silent green skip. There is also no dedicated CI job
+that names raft e2e as a first-class gate.
+
+### G3. Cold node — two real failure modes
+
+1. **Log growth pinning.** Leader compaction
+   (`raft.nim:244-276`, `compactLog`) never discards past any peer's
+   `matchIndex`. A peer that is down keeps `matchIndex` stale, so the leader's
+   log grows without bound for as long as the node is down.
+2. **Unrecoverable laggard.** Once the leader's log no longer contains a
+   follower's `nextIndex` (fresh/wiped node, or a node that was down through a
+   compaction), the follower rejects every AppendEntries
+   (`handleAppendEntries`, `raft.nim:380-394`) and the leader's
+   `nextIndex` decrement floor is `lastSnapshotIndex + 1`
+   (`handleAppendReply`, `raft.nim:526-531`). The pair is stuck forever: no
+   InstallSnapshot path exists.
+
+### G4. Raft port is plaintext
+
+`RaftNetwork` uses bare `newAsyncSocket()` (`raft.nim:748-765`, `857-871`).
+Any host that can reach the raft port can inject RequestVote/AppendEntries
+frames. The TLS machinery in `protocol/ssl.nim` is not used here; leader
+forwarding (`forwardQueryToLeader`) is likewise plaintext.
+
+### G5. Raft write encoding loses empty-value puts (found 2026-07-30)
+
+`appendWriteToRaft` (`core/server.nim:309-330`) encodes an empty value as
+`"delete"`, but PK-only-table inserts legitimately produce empty values
+(`execInsert`, `query/exec/dml.nim:60-90`) — such inserts are acked after
+majority commit and then deleted everywhere on apply. Fixed as plan Task 1a
+(explicit `deleted` flag on `ExecResult.keyValuePairs`).
+
+## Design
+
+### D1. Failover-under-load E2E (test-only)
+
+New suite `tests/raft_failover_load_e2e_test.nim`, same process-management
+conventions as `raft_writes_e2e_test.nim` (port base `50000 + tstamp mod
+4000` to avoid collisions):
+
+1. Boot a 3-node cluster, elect a leader, create `load_test` table via raft
+   DDL.
+2. Writer thread: sequential `INSERT INTO load_test (id) VALUES (n)`,
+   `n = 1, 2, ...`, recording every *acknowledged* id. On error ("not
+   leader", commit timeout, connection reset), probe both survivors and
+   resume on whichever accepts.
+3. At ~50 acknowledged writes, kill the leader.
+4. Assert: a survivor accepts a write within **10 s** of the kill
+   (availability bound).
+5. Assert: after the new leader is stable and the remaining follower has
+   caught up, `SELECT id FROM load_test` on **both** survivors contains
+   **every acknowledged id** (committed writes never lost). Unacknowledged
+   writes may be present or absent — this is documented, not asserted.
+
+Also document the client-visible contract in `docs/en/distributed.md`:
+in-flight writes during failover fail fast with an error; clients must
+retry; acknowledged writes are durable across failover.
+
+### D2. CI e2e mandatory
+
+- New `raft-e2e` job in `.github/workflows/ci.yml`: setup Nim, build
+  `build/baradadb`, run the three raft e2e suites explicitly with `CI=true`
+  in the environment.
+- Change skip semantics in all raft e2e suites: when `CI` env var is
+  non-empty and `./build/baradadb` is missing, **fail** instead of `skip()`.
+- Add `raft_failover_load_e2e_test` to the `nimble test` list in
+  `baradadb.nimble`.
+
+### D3. Cold node — InstallSnapshot
+
+Extend the raft wire protocol and apply path:
+
+**Protocol.** New message kinds `rmkInstallSnapshot`,
+`rmkInstallSnapshotReply`, and new fields on `RaftMessage`:
+`snapId: uint64`, `snapOffset: uint64`, `snapData: seq[byte]`,
+`snapDone: bool`. `lastSnapshotIndex`/`lastSnapshotTerm` ride on the
+existing fields (`prevLogIndex`/`prevLogTerm` are reused as the snapshot
+base for this kind). Serialization appends the new fields with `atEnd`
+guards (same backward-compatible pattern as `loadState`,
+`raft.nim:163-167`); `RaftProtoVersion` stays 1 — mixed-version clusters
+simply never send the new kind (old leaders never trigger it).
+
+**Leader side.** Track consecutive AppendEntries rejections per peer. When
+`nextIndex[peer]` has hit the `lastSnapshotIndex + 1` floor and the peer
+still rejects, the peer is unrecoverably behind:
+
+1. Build a snapshot archive of the **default database** data dir with the
+   existing backup machinery (`backupDataDir` in
+   `src/barabadb/core/backup.nim:225`) into a temp file.
+2. Stream it in chunks (`BARADB_RAFT_SNAP_CHUNK_KB`, default 256 KB) as
+   `rmkInstallSnapshot` messages over the existing peer socket.
+3. On final ack, set `matchIndex[peer] = lastSnapshotIndex`,
+   `nextIndex[peer] = lastSnapshotIndex + 1`, resume normal AppendEntries.
+
+**Follower side.** On `rmkInstallSnapshot`:
+
+1. Buffer chunks to a temp file under `dataDir/raft/snap_incoming/`.
+2. On `snapDone`, hand the archive to a new injected callback
+   `restoreSnapshot: proc(archivePath: string): bool {.gcsafe.}` (wired in
+   `baradadb.nim` where the `DatabaseRegistry` lives): close the default
+   DB, `restoreDataDir` (`backup.nim:263`) into the default DB dir, reopen,
+   and swap execution context.
+3. Set `lastSnapshotIndex`/`lastSnapshotTerm`/`commitIndex`/`lastApplied`
+   from the message, clear the log, `saveState`.
+
+**Unpinning compaction.** Once InstallSnapshot exists, `compactLog` on the
+leader compacts through `lastApplied` for peers whose `matchIndex` was
+updated within the last `BARADB_RAFT_PEER_STALE_MS` (default 30 000);
+long-dead peers no longer pin the log — they get a snapshot when they
+return. Follower compaction is unchanged.
+
+**Fresh-node join** falls out for free: a wiped node rejects at the floor
+and receives a snapshot.
+
+### D4. Raft TLS
+
+Config (env, mirroring existing `BARADB_TLS_*`):
+
+| Env | Config field | Default |
+|-----|--------------|---------|
+| `BARADB_RAFT_TLS_ENABLED` | `raftTlsEnabled: bool` | false |
+| `BARADB_RAFT_TLS_CERT_FILE` | `raftTlsCertFile: string` | "" |
+| `BARADB_RAFT_TLS_KEY_FILE` | `raftTlsKeyFile: string` | "" |
+| `BARADB_RAFT_TLS_CA_FILE` | `raftTlsCaFile: string` | "" |
+| `BARADB_RAFT_TLS_VERIFY_PEER` | `raftTlsVerifyPeer: bool` | false |
+
+- `RaftNetwork` gains `tls: TLSContext` (nil = plaintext, current
+  behavior). `connectToPeer` wraps with `wrapClient`; the accept loop in
+  `run` wraps with `wrapServer` before `receiveLoop` (same pattern as
+  `core/server.nim:876-889`). `verifyPeer` + CA file gives mutual auth.
+- Fail closed: `raftEnabled and raftTlsEnabled` with missing cert/key →
+  refuse to start (raise at startup, like the JWT check in
+  `newServerWithRegistry`, `core/server.nim:58-63`).
+- Leader SQL forwarding (`forwardQueryToLeader`) wraps its socket with the
+  **client** TLS context when `BARADB_TLS_ENABLED` is on (it dials the
+  client wire port, which is already TLS-capable).
+- E2E: TLS variant cluster test — generate self-signed certs with
+  `generateSelfSignedCert` (`protocol/ssl.nim:79`), boot a 3-node cluster
+  with raft TLS on, assert election + one replicated write; assert a
+  plaintext peer cannot join (its frames are rejected and the cluster
+  elects among the TLS nodes).
+
+## Rollout / phases
+
+| Phase | Deliverable | Risk |
+|-------|-------------|------|
+| P1 | D1 failover-load e2e | none (test-only) |
+| P2 | D2 CI e2e job | none |
+| P3 | D4 raft TLS | medium (transport) |
+| P4 | D3 InstallSnapshot + compaction unpin | high (protocol + apply) |
+
+P3 before P4 so snapshot transfer ships already-encryptable. Each phase is
+independently mergeable; P4 is the v1.3.0 gate for calling raft
+"supported".
+
+## Acceptance (v1.3.0)
+
+- Failover-under-load e2e green locally and in CI, in the mandatory gate.
+- Killed-node-returns and wiped-node-join scenarios converge without manual
+  intervention (covered by new e2e phases).
+- Leader log length stays bounded while a peer is down > `PEER_STALE_MS`
+  (assert via `baradb_raft_log_entries` metric in e2e).
+- Raft port TLS: cluster runs fully over TLS; plaintext injection fails.
+- `docs/en/distributed.md` + `known-limitations.md` updated: raft no longer
+  "experimental" for the covered scope; remaining non-goals listed.

src/barabadb/core/backup.nim

@@ -26,6 +26,8 @@ import std/strutils
 import std/times
 import std/algorithm
 import std/json
+import std/asyncdispatch
+import std/threadpool
 import barabadb/storage/lsm
 
 type
@@ -260,6 +262,80 @@ proc backupDataDir*(dataDir: string, output: string, excludes: seq[string] = @[]
   echo "  Source:   ", dataDir
   return true
 
+proc tarDataDir*(dataDir: string, output: string, excludes: seq[string] = @[]): bool =
+  ## Create an UNCOMPRESSED tar of `dataDir` at `output` (no gzip). The raft
+  ## snapshot sender runs this under the storage gate for a consistent file
+  ## capture, then compresses off the event loop via gzipFileAsync.
+  if not dirExists(dataDir):
+    echo "ERROR: Data directory not found: ", dataDir
+    return false
+
+  let parent = parentDir(dataDir)
+  let name = lastPathPart(dataDir)
+  var excludeArgs = ""
+  for pattern in excludes:
+    excludeArgs.add(" --exclude=" & quoteShell(pattern))
+
+  let cmd = "tar -cf " & quoteShell(output) & excludeArgs &
+            " -C " & quoteShell(parent) & " " & quoteShell(name)
+  let (outputStr, exitCode) = execCmdEx(cmd)
+  if exitCode != 0:
+    echo "ERROR: tar command failed with exit code ", exitCode
+    if outputStr.len > 0:
+      echo outputStr
+    return false
+  return true
+
+proc gzipFile*(input: string, output: string,
+               compression: int = DEFAULT_COMPRESSION): bool =
+  ## gzip a single file `input` -> `output`. Pure CPU over an already-captured
+  ## file: no shared storage state, so it is safe to run on a worker thread
+  ## outside the storage gate and off the raft event loop.
+  if not fileExists(input):
+    echo "ERROR: File not found: ", input
+    return false
+
+  let cmd = "gzip -" & $compression & " -c " & quoteShell(input) &
+            " > " & quoteShell(output)
+  let (outputStr, exitCode) = execCmdEx("bash -c " & quoteShell(cmd))
+  if exitCode != 0:
+    echo "ERROR: gzip command failed with exit code ", exitCode
+    if outputStr.len > 0:
+      echo outputStr
+    return false
+  return true
+
+proc gunzipFile*(input: string, output: string): bool =
+  ## Decompress a gzip file `input` -> `output`. Inverse of gzipFile.
+  if not fileExists(input):
+    echo "ERROR: File not found: ", input
+    return false
+
+  let cmd = "gzip -dc " & quoteShell(input) & " > " & quoteShell(output)
+  let (outputStr, exitCode) = execCmdEx("bash -c " & quoteShell(cmd))
+  if exitCode != 0:
+    echo "ERROR: gunzip command failed with exit code ", exitCode
+    if outputStr.len > 0:
+      echo outputStr
+    return false
+  return true
+
+proc gzipFileWorker(input: string, output: string, compression: int): bool {.gcsafe.} =
+  ## Thread entry point: touches only its own (copied) string args + execCmdEx,
+  ## so it is safe to run off the main/event-loop thread under ARC/ORC.
+  gzipFile(input, output, compression)
+
+proc gzipFileAsync*(input: string, output: string,
+                    compression: int = DEFAULT_COMPRESSION): Future[bool] {.async.} =
+  ## Run gzipFile on a threadpool worker and await completion WITHOUT blocking
+  ## the calling async event loop — heartbeats/election timers keep firing
+  ## during the CPU-heavy compression. Polls the FlowVar via sleepAsync so the
+  ## dispatcher stays responsive instead of stalling on a blocking join.
+  var fv = spawn gzipFileWorker(input, output, compression)
+  while not fv.isReady:
+    await sleepAsync(20)
+  result = ^fv
+
 proc restoreDataDir*(input: string, dataDir: string, verbose: bool = false, dryRun: bool = false): bool =
   ## Restore from a tar.gz backup.
   ## When dryRun is true, only prints what would be done.

src/barabadb/core/config.nim

@@ -14,6 +14,8 @@ type
     tlsEnabled*: bool
     certFile*: string
     keyFile*: string
+    tlsCaFile*: string
+    tlsVerifyPeer*: bool
     idleTimeoutMs*: int
     queryTimeoutMs*: int
     slowQueryThresholdMs*: int
@@ -44,6 +46,13 @@ type
     raftPeerClientAddrs*: Table[string, tuple[host: string, port: int]]
     raftWriteTimeoutMs*: int
     raftLogMaxEntries*: int
+    raftSnapChunkKb*: int
+    raftPeerStaleMs*: int
+    raftTlsEnabled*: bool
+    raftTlsCertFile*: string
+    raftTlsKeyFile*: string
+    raftTlsCaFile*: string
+    raftTlsVerifyPeer*: bool
 
   CompactionStrategy* = enum
     csSizeTiered = "size_tiered"
@@ -60,6 +69,8 @@ proc defaultConfig*(): BaraConfig =
     tlsEnabled: false,
     certFile: "",
     keyFile: "",
+    tlsCaFile: "",
+    tlsVerifyPeer: false,
     idleTimeoutMs: 300_000,
     queryTimeoutMs: 30_000,
     slowQueryThresholdMs: 1_000,
@@ -86,6 +97,13 @@ proc defaultConfig*(): BaraConfig =
     raftPeerClientAddrs: initTable[string, tuple[host: string, port: int]](),
     raftWriteTimeoutMs: 5_000,
     raftLogMaxEntries: 256,
+    raftSnapChunkKb: 256,
+    raftPeerStaleMs: 30000,
+    raftTlsEnabled: false,
+    raftTlsCertFile: "",
+    raftTlsKeyFile: "",
+    raftTlsCaFile: "",
+    raftTlsVerifyPeer: false,
   )
 
 # ----------------------------------------------------------------------
@@ -120,6 +138,8 @@ proc loadConfigFromJson*(path: string, cfg: var BaraConfig) =
       if s.hasKey("enabled"): cfg.tlsEnabled = s["enabled"].getBool()
       if s.hasKey("cert_file"): cfg.certFile = s["cert_file"].getStr()
       if s.hasKey("key_file"): cfg.keyFile = s["key_file"].getStr()
+      if s.hasKey("ca_file"): cfg.tlsCaFile = s["ca_file"].getStr()
+      if s.hasKey("verify_peer"): cfg.tlsVerifyPeer = s["verify_peer"].getBool()
     if j.hasKey("auth"):
       let s = j["auth"]
       if s.hasKey("enabled"): cfg.authEnabled = s["enabled"].getBool()
@@ -163,6 +183,13 @@ proc loadConfigFromEnv*(cfg: var BaraConfig) =
   cfg.tlsEnabled = parseEnvBool(getEnv("BARADB_TLS_ENABLED", ""), cfg.tlsEnabled)
   cfg.certFile = getEnv("BARADB_CERT_FILE", cfg.certFile)
   cfg.keyFile = getEnv("BARADB_KEY_FILE", cfg.keyFile)
+  cfg.tlsCaFile = getEnv("BARADB_TLS_CA_FILE", cfg.tlsCaFile)
+  let tlsVerifyEnv = getEnv("BARADB_TLS_VERIFY_PEER", "")
+  if tlsVerifyEnv.len > 0:
+    cfg.tlsVerifyPeer = parseEnvBool(tlsVerifyEnv, cfg.tlsVerifyPeer)
+  elif cfg.tlsCaFile.len > 0:
+    # CA present and verify flag unset → verify (fail-closed for MITM).
+    cfg.tlsVerifyPeer = true
   cfg.idleTimeoutMs = parseEnvInt(getEnv("BARADB_IDLE_TIMEOUT_MS", ""), cfg.idleTimeoutMs)
   cfg.queryTimeoutMs = parseEnvInt(getEnv("BARADB_QUERY_TIMEOUT_MS", ""), cfg.queryTimeoutMs)
   cfg.slowQueryThresholdMs = parseEnvInt(getEnv("BARADB_SLOW_QUERY_THRESHOLD_MS", ""), cfg.slowQueryThresholdMs)
@@ -213,6 +240,17 @@ proc loadConfigFromEnv*(cfg: var BaraConfig) =
   cfg.raftNodeId = getEnv("BARADB_RAFT_NODE_ID", cfg.raftNodeId)
   cfg.raftWriteTimeoutMs = parseEnvInt(getEnv("BARADB_RAFT_WRITE_TIMEOUT_MS", ""), cfg.raftWriteTimeoutMs)
   cfg.raftLogMaxEntries = parseEnvInt(getEnv("BARADB_RAFT_LOG_MAX_ENTRIES", ""), cfg.raftLogMaxEntries)
+  cfg.raftSnapChunkKb = parseEnvInt(getEnv("BARADB_RAFT_SNAP_CHUNK_KB", ""), cfg.raftSnapChunkKb)
+  cfg.raftPeerStaleMs = parseEnvInt(getEnv("BARADB_RAFT_PEER_STALE_MS", ""), cfg.raftPeerStaleMs)
+  cfg.raftTlsEnabled = parseEnvBool(getEnv("BARADB_RAFT_TLS_ENABLED", ""), cfg.raftTlsEnabled)
+  cfg.raftTlsCertFile = getEnv("BARADB_RAFT_TLS_CERT_FILE", cfg.raftTlsCertFile)
+  cfg.raftTlsKeyFile = getEnv("BARADB_RAFT_TLS_KEY_FILE", cfg.raftTlsKeyFile)
+  cfg.raftTlsCaFile = getEnv("BARADB_RAFT_TLS_CA_FILE", cfg.raftTlsCaFile)
+  let raftVerifyEnv = getEnv("BARADB_RAFT_TLS_VERIFY_PEER", "")
+  if raftVerifyEnv.len > 0:
+    cfg.raftTlsVerifyPeer = parseEnvBool(raftVerifyEnv, cfg.raftTlsVerifyPeer)
+  elif cfg.raftTlsCaFile.len > 0:
+    cfg.raftTlsVerifyPeer = true
   # Optional: client (SQL) addresses for leader write forwarding.
   # Same id@host:port shape as BARADB_RAFT_PEERS, but ports are BARADB_PORT values.
   let clientPeersEnv = getEnv("BARADB_RAFT_CLIENT_PEERS", "")
@@ -271,6 +309,14 @@ proc validateProductionConfig*(cfg: BaraConfig) =
   if cfg.jwtSecret in ["change-me", "change-me-to-random-32-char-string", "secret", "default"]:
     raise newException(ValueError,
       "Production refuses insecure JWT secret placeholder. Set a strong BARADB_JWT_SECRET.")
+  if cfg.tlsEnabled:
+    if not cfg.tlsVerifyPeer or cfg.tlsCaFile.len == 0:
+      raise newException(ValueError,
+        "Production TLS requires peer verification. Set BARADB_TLS_VERIFY_PEER=true and BARADB_TLS_CA_FILE.")
+  if cfg.raftTlsEnabled:
+    if not cfg.raftTlsVerifyPeer or cfg.raftTlsCaFile.len == 0:
+      raise newException(ValueError,
+        "Production raft TLS requires peer verification. Set BARADB_RAFT_TLS_VERIFY_PEER=true and BARADB_RAFT_TLS_CA_FILE.")
 
 proc getEffectiveJwtSecret*(cfg: BaraConfig): string =
   if cfg.jwtSecret.len > 0:

src/barabadb/core/disttxn.nim

@@ -5,6 +5,8 @@ import std/monotimes
 import std/net
 import std/strutils
 import std/nativesockets
+when defined(posix):
+  import std/posix
 
 type
   DistTxnState* = enum
@@ -89,6 +91,13 @@ proc connectWithTimeout(sock: Socket, host: string, port: Port, timeoutMs: int):
     var fds = @[sock.getFd]
     if selectWrite(fds, timeoutMs) <= 0:
       return false
+    when defined(posix):
+      # selectWrite reports a refused connect as writable; SO_ERROR tells the truth.
+      var err: cint = 0
+      var errLen = SockLen(sizeof(err))
+      discard posix.getsockopt(sock.getFd, 1'i32, 4'i32, addr err, addr errLen)
+      if err != 0:
+        return false
     sock.getFd.setBlocking(true)
     return true
 
@@ -96,15 +105,18 @@ proc sendDistTxnRpc(host: string, port: int, txnId: uint64, action: string, time
   ## Send 2PC RPC to participant node via TCP text protocol.
   ## Protocol: "DISTTXN <txnId> <action>\n" where action = PREPARE|COMMIT|ROLLBACK
   ## Response: "OK\n" or "ERR <msg>\n"
-  var sock = newSocket()
-  defer: sock.close()
-  if not connectWithTimeout(sock, host, Port(port), timeoutMs):
+  try:
+    var sock = newSocket()
+    defer: sock.close()
+    if not connectWithTimeout(sock, host, Port(port), timeoutMs):
+      return false
+    let msg = "DISTTXN " & $txnId & " " & action & "\n"
+    sock.send(msg)
+    var response = ""
+    sock.readLine(response)
+    return response.strip() == "OK"
+  except CatchableError:
     return false
-  let msg = "DISTTXN " & $txnId & " " & action & "\n"
-  sock.send(msg)
-  var response = ""
-  sock.readLine(response)
-  return response.strip() == "OK"
 
 type
   ParticipantInfo = object

src/barabadb/core/httpserver.nim

@@ -19,6 +19,7 @@ import ../storage/gate
 import ../core/mvcc
 import ../protocol/wire
 import ../core/websocket
+import ../query/exec/rls
 import jwt as jwtlib
 import ../protocol/auth
 import ../protocol/ratelimit
@@ -31,7 +32,7 @@ type
     config: BaraConfig
     running: bool
     db*: LSMTree
-    ctx: ExecutionContext
+    ctx*: ExecutionContext  # read/write only under the storage gate
     registry*: DatabaseRegistry
     metrics*: Metrics
     secretKey*: string
@@ -55,6 +56,14 @@ proc newHttpServerWithRegistry*(config: BaraConfig, registry: DatabaseRegistry):
   ctx.txnManager = newTxnManager()
   let secret = config.getEffectiveJwtSecret()
   let ws = newWsServer(config, secret)
+  block:
+    let wsRef {.cursor.} = ws
+    let ctxRef {.cursor.} = ctx
+    wsRef.canSubscribe = proc(username, table: string): bool {.gcsafe.} =
+      if username.len == 0:
+        return false
+      {.cast(gcsafe).}:
+        return hasPrivilegeFor(ctxRef, username, table, "SELECT")
   let rl = newRateLimiter(rlaTokenBucket, config.rateLimitGlobal, config.rateLimitPerClient)
   ctx.onChange = proc(ev: ChangeEvent) =
     let msg = $ev.kind & " " & ev.table
@@ -262,7 +271,7 @@ proc healthHandler(server: HttpServer): RequestHandler =
     let ctx = newContext(request)
     var body = %*{
       "status": "ok",
-      "version": "1.2.0"
+      "version": "1.3.0"
     }
     if server.raftNode != nil:
       let n = server.raftNode
@@ -368,7 +377,7 @@ proc openApiHandler(): RequestHandler =
     let ctx = newContext(request)
     ctx.json(%*{
       "openapi": "3.0.0",
-      "info": {"title": "BaraDB API", "version": "1.2.0"},
+      "info": {"title": "BaraDB API", "version": "1.3.0"},
       "paths": {
         "/query": {
           "post": {
@@ -906,7 +915,7 @@ function showTab(idx){
 }
 setInterval(() => { if(document.querySelectorAll('.panel')[4].classList.contains('active')) loadMetrics() }, 5000)
 </script>
-<div class='status' style='text-align:center;padding:10px'>BaraDB v1.2.0 — Multimodal Database Engine</div>
+<div class='status' style='text-align:center;padding:10px'>BaraDB v1.3.0 — Multimodal Database Engine</div>
 </body></html>"""
     request.respond(200, @[("Content-Type", "text/html; charset=utf-8")], html)
 

src/barabadb/core/mvcc.nim

@@ -178,12 +178,16 @@ proc write*(tm: TxnManager, txn: Transaction, key: string, value: seq[byte]): bo
     return false
 
   # Timeout-based deadlock detection: abort stale transactions
+  # Collect then delete — never mutate activeTxns while iterating it.
   let now = getMonoTime().ticks()
+  var staleIds: seq[TxnId] = @[]
   for otherId, otherTxn in tm.activeTxns:
     if otherId != txn.id and otherTxn.state == tsActive:
       if now - otherTxn.startTime > tm.txnTimeoutMs * 1_000_000:
         otherTxn.state = tsAborted
-        tm.activeTxns.del(otherId)
+        staleIds.add(otherId)
+  for id in staleIds:
+    tm.activeTxns.del(id)
 
   # Check for write-write conflict against other active transactions' write sets
   for otherId, otherTxn in tm.activeTxns:

src/barabadb/core/raft.nim

@@ -12,6 +12,8 @@ import std/endians
 import std/os
 import logging
 import ../protocol/wire
+import ../protocol/ssl
+import backup
 
 type
   RaftState* = enum
@@ -66,6 +68,15 @@ type
     # Leader state
     nextIndex*: Table[string, uint64]
     matchIndex*: Table[string, uint64]
+    ## Monotonic ms of the last successful AppendEntries reply per peer,
+    ## initialized to "now" in becomeLeader (grace window) and bumped in
+    ## handleAppendReply. Leader compaction excludes peers silent longer than
+    ## raftPeerStaleMs from its minMatch — a stale peer no longer pins the log
+    ## forever; it is caught up via InstallSnapshot on return (T9).
+    matchIndexSeenMs*: Table[string, int64]
+    ## Stale window in ms (BARADB_RAFT_PEER_STALE_MS, default 30000;
+    ## 0 = default).
+    raftPeerStaleMs*: int
     # Cluster
     peers*: seq[string]
     leaderId*: string
@@ -76,12 +87,35 @@ type
     peerAddrs*: Table[string, tuple[host: string, port: int]]
     raftPort*: int
     dataDir*: string
+    ## InstallSnapshot follower receive. snapChunkBytes caps a single chunk
+    ## (from BARADB_RAFT_SNAP_CHUNK_KB, default 262144); snapIncomingId /
+    ## snapIncomingFile track the archive currently being assembled under
+    ## dataDir/snap_incoming/.
+    snapChunkBytes*: int
+    restoreSnapshot*: proc(archivePath: string, baseIndex: uint64,
+                           baseTerm: uint64): bool {.gcsafe.}
+    snapIncomingId*: uint64
+    snapIncomingFile*: string
+    ## Leader InstallSnapshot send. buildSnapshot writes an UNCOMPRESSED tar
+    ## of the data dir to destPath (wired in baradadb.nim via tarDataDir, under
+    ## the storage gate); sendSnapshot then compresses it off the event loop
+    ## (gzipFileAsync) and streams the resulting .tar.gz to the follower.
+    ## snapRejectStreak counts consecutive floor-level AppendEntries rejects
+    ## per peer; at 2 the peer is queued in snapPending and the network layer
+    ## (processMessage) kicks off sendSnapshot. snapSending is the
+    ## single-flight guard: at most one snapshot transfer per peer.
+    buildSnapshot*: proc(destPath: string): bool {.gcsafe.}
+    snapRejectStreak*: Table[string, int]
+    snapPending*: HashSet[string]
+    snapSending*: HashSet[string]
 
   RaftMessageKind* = enum
     rmkRequestVote
     rmkRequestVoteReply
     rmkAppendEntries
     rmkAppendEntriesReply
+    rmkInstallSnapshot
+    rmkInstallSnapshotReply
 
   RaftMessage* = object
     kind*: RaftMessageKind
@@ -98,6 +132,12 @@ type
     # Reply
     success*: bool
     matchIdx*: uint64
+    # InstallSnapshot (prevLogIndex/prevLogTerm reuse: snapshot base index/term;
+    # reply uses success/matchIdx as usual)
+    snapId*: uint64      # snapshot generation, matches leader's base at build time
+    snapOffset*: uint64  # byte offset of this chunk within the archive
+    snapData*: seq[byte] # chunk payload (<= snapChunkBytes)
+    snapDone*: bool      # last chunk
 
   RaftCluster* = ref object
     nodes*: Table[string, RaftNode]
@@ -191,6 +231,8 @@ proc newRaftNode*(id: string, peers: seq[string], raftPort: int = 0,
     metrics: RaftMetrics(),
     nextIndex: initTable[string, uint64](),
     matchIndex: initTable[string, uint64](),
+    matchIndexSeenMs: initTable[string, int64](),
+    raftPeerStaleMs: 30000,
     peers: peers,
     leaderId: "",
     electionTimeout: 150 + rand(150),
@@ -199,6 +241,12 @@ proc newRaftNode*(id: string, peers: seq[string], raftPort: int = 0,
     peerAddrs: initTable[string, tuple[host: string, port: int]](),
     raftPort: raftPort,
     dataDir: dataDir,
+    snapChunkBytes: 262144,
+    snapIncomingId: 0,
+    snapIncomingFile: "",
+    snapRejectStreak: initTable[string, int](),
+    snapPending: initHashSet[string](),
+    snapSending: initHashSet[string](),
   )
   result.loadState()
 
@@ -243,15 +291,22 @@ proc termAtIndex(node: RaftNode, index: uint64): uint64 =
 
 proc compactLog*(node: RaftNode) =
   ## Drop a fully-replicated / applied log prefix so the in-memory log stays
-  ## bounded. Leader: never discard past any peer's matchIndex (catch-up via
-  ## AppendEntries remains possible). Follower: discard through lastApplied.
+  ## bounded. Leader: never discard past any responsive peer's matchIndex
+  ## (catch-up via AppendEntries remains possible); peers silent longer than
+  ## raftPeerStaleMs are excluded and catch up via InstallSnapshot instead.
+  ## Follower: discard through lastApplied.
   let maxEntries = if node.logMaxEntries > 0: node.logMaxEntries else: 256
   if node.log.len <= maxEntries:
     return
   var through = node.lastApplied
   if node.state == rsLeader and node.peers.len > 0:
+    let staleMs = if node.raftPeerStaleMs > 0: node.raftPeerStaleMs else: 30000
+    let nowMs = getMonoTime().ticks() div 1_000_000
     var minMatch = through
     for peer in node.peers:
+      let seenMs = node.matchIndexSeenMs.getOrDefault(peer, 0)
+      if seenMs <= 0 or nowMs - seenMs > staleMs.int64:
+        continue  # stale peer — unpinned, snapshotted on return (T9)
       let m = node.matchIndex.getOrDefault(peer, 0'u64)
       if m < minMatch: minMatch = m
     through = minMatch
@@ -312,6 +367,10 @@ proc becomeFollower*(node: RaftNode, term: uint64) =
   node.votesReceived.clear()
   node.nextIndex.clear()
   node.matchIndex.clear()
+  node.matchIndexSeenMs.clear()
+  # Leader-only snapshot-send state is meaningless once we step down
+  node.snapRejectStreak.clear()
+  node.snapPending.clear()
   node.saveState()
 
 proc becomeCandidate*(node: RaftNode) =
@@ -330,9 +389,16 @@ proc becomeLeader*(node: RaftNode) =
   if node.metrics != nil:
     inc node.metrics.electionsTotal
   info("Raft node " & node.id & " became leader for term " & $node.currentTerm)
+  let nowMs = getMonoTime().ticks() div 1_000_000
+  node.matchIndexSeenMs.clear()
   for peer in node.peers:
     node.nextIndex[peer] = node.lastLogIndex + 1
     node.matchIndex[peer] = 0
+    # Grace window: an unreplied peer still pins compaction until it has been
+    # silent for raftPeerStaleMs since this leadership began.
+    node.matchIndexSeenMs[peer] = nowMs
+  node.snapRejectStreak.clear()
+  node.snapPending.clear()
 
 proc handleRequestVote*(node: RaftNode, msg: RaftMessage): RaftMessage =
   var reply = RaftMessage(
@@ -418,6 +484,87 @@ proc handleAppendEntries*(node: RaftNode, msg: RaftMessage): RaftMessage =
   reply.matchIdx = node.lastLogIndex
   return reply
 
+proc handleInstallSnapshot*(node: RaftNode, msg: RaftMessage): RaftMessage =
+  ## Follower side of InstallSnapshot: assemble the chunk stream into a temp
+  ## archive under `dataDir/snap_incoming/`, then hand the completed archive
+  ## to the restoreSnapshot callback. Chunks arrive in order from a single
+  ## leader over one socket, so we append sequentially and only sanity-check
+  ## that snapOffset equals the number of bytes assembled so far.
+  ##
+  ## NOTE: this runs on the async event loop and restoreSnapshot performs
+  ## blocking disk I/O (archive extract + DB reopen). Implementations must be
+  ## fast, or defer the heavy work; the baradadb.nim wiring decides.
+  var reply = RaftMessage(
+    kind: rmkInstallSnapshotReply,
+    term: node.currentTerm,
+    senderId: node.id,
+    success: false,
+    matchIdx: node.lastSnapshotIndex,
+  )
+  if msg.term < node.currentTerm:
+    return reply
+  if msg.term > node.currentTerm:
+    node.becomeFollower(msg.term)
+  node.leaderId = msg.senderId
+
+  # Chunk size cap (deferred from the wire-protocol task).
+  if msg.snapData.len > node.snapChunkBytes or node.dataDir.len == 0:
+    return reply
+
+  let snapDir = node.dataDir / "snap_incoming"
+  if msg.snapId != node.snapIncomingId:
+    # New snapshot generation: discard any partial assembly and restart.
+    if msg.snapOffset != 0:
+      return reply
+    createDir(snapDir)
+    node.snapIncomingId = msg.snapId
+    node.snapIncomingFile = snapDir / "snap_" & $msg.snapId & ".tar.gz"
+    let f = open(node.snapIncomingFile, fmWrite)  # truncate any leftover
+    f.close()
+
+  if node.snapIncomingFile.len == 0:
+    return reply
+
+  let assembled = getFileSize(node.snapIncomingFile)
+  if msg.snapOffset != uint64(assembled):
+    # Gap or overlap: reset so the leader restarts the transfer.
+    removeFile(node.snapIncomingFile)
+    node.snapIncomingId = 0
+    node.snapIncomingFile = ""
+    return reply
+
+  if msg.snapData.len > 0:
+    let f = open(node.snapIncomingFile, fmAppend)
+    try:
+      discard f.writeBuffer(addr msg.snapData[0], msg.snapData.len)
+    finally:
+      f.close()
+
+  if not msg.snapDone:
+    reply.success = true
+    return reply
+
+  # Transfer complete: restore the data dir and adopt the snapshot base.
+  if node.restoreSnapshot == nil or
+      not node.restoreSnapshot(node.snapIncomingFile,
+                               msg.prevLogIndex, msg.prevLogTerm):
+    removeFile(node.snapIncomingFile)
+    node.snapIncomingId = 0
+    node.snapIncomingFile = ""
+    return reply
+
+  node.lastSnapshotIndex = msg.prevLogIndex
+  node.lastSnapshotTerm = msg.prevLogTerm
+  node.commitIndex = node.lastSnapshotIndex
+  node.lastApplied = node.lastSnapshotIndex
+  node.log = @[]
+  node.snapIncomingId = 0
+  node.snapIncomingFile = ""
+  node.saveState()
+  reply.success = true
+  reply.matchIdx = node.lastSnapshotIndex
+  return reply
+
 proc requestVote*(node: RaftNode): seq[RaftMessage] =
   result = @[]
   for peer in node.peers:
@@ -498,9 +645,15 @@ proc handleAppendReply*(node: RaftNode, peerId: string, reply: RaftMessage) =
   if reply.success:
     node.matchIndex[peerId] = reply.matchIdx
     node.nextIndex[peerId] = reply.matchIdx + 1
+    node.matchIndexSeenMs[peerId] = getMonoTime().ticks() div 1_000_000
+    node.snapRejectStreak.del(peerId)
+    node.snapPending.excl(peerId)
 
-    # Update commit index using true majority calculation
-    let majority = (node.peers.len + 1 + 1) div 2  # majority of cluster (peers + leader)
+    # Update commit index using strict majority — the same form as the election
+    # check in handleVoteReply. Cluster size N = peers.len + 1; a strict
+    # majority is N div 2 + 1. The previous (N + 1) div 2 under-counted for
+    # even-sized clusters (e.g. N=4 committed at 2/4, a minority).
+    let majority = (node.peers.len + 1) div 2 + 1  # strict majority of cluster
     var newCommitIdx = node.commitIndex
 
     # Walk logical indices high→low via findLogEntryByIndex (log may be compacted).
@@ -527,8 +680,47 @@ proc handleAppendReply*(node: RaftNode, peerId: string, reply: RaftMessage) =
     let floor = node.lastSnapshotIndex + 1
     if node.nextIndex.getOrDefault(peerId, 1) > floor:
       dec node.nextIndex[peerId]
+      # Not a floor-level reject, so it breaks any floor-reject streak.
+      node.snapRejectStreak.del(peerId)
     else:
       node.nextIndex[peerId] = floor
+      # Stuck at the compaction floor: the entries the follower needs have
+      # been compacted away, so AppendEntries can never catch it up. Count
+      # consecutive floor rejects; at 2, queue an InstallSnapshot transfer
+      # (the network layer picks this up after handleAppendReply returns).
+      node.snapRejectStreak[peerId] =
+        node.snapRejectStreak.getOrDefault(peerId, 0) + 1
+      if node.snapRejectStreak[peerId] >= 2:
+        node.snapPending.incl(peerId)
+
+proc handleInstallSnapshotReply*(node: RaftNode, peerId: string,
+                                 reply: RaftMessage) =
+  ## Leader side: follower's answer to a completed InstallSnapshot transfer.
+  ## success=true adopts the snapshot base (reply.matchIdx) as the peer's
+  ## match point; success=false leaves all state alone — the normal
+  ## AppendEntries reject path re-triggers another snapshot if the peer is
+  ## still stuck at the floor.
+  if reply.term > node.currentTerm:
+    node.becomeFollower(reply.term)
+    return
+
+  if reply.term < node.currentTerm:
+    return
+
+  if node.state != rsLeader:
+    return
+
+  if reply.success and reply.matchIdx >= node.lastSnapshotIndex:
+    # The follower has actually adopted the snapshot base. Intermediate chunk
+    # replies (the T8 follower acks every non-done chunk with success=true and
+    # matchIdx = its OLD lastSnapshotIndex, below ours) fall through here and
+    # must be ignored: applying them would regress matchIndex/nextIndex and
+    # clear the reject streak mid-transfer, causing state flapping until the
+    # final reply lands.
+    node.matchIndex[peerId] = reply.matchIdx
+    node.nextIndex[peerId] = reply.matchIdx + 1
+    node.snapRejectStreak.del(peerId)
+    node.snapPending.excl(peerId)
 
 proc state*(node: RaftNode): RaftState = node.state
 proc isLeader*(node: RaftNode): bool = node.state == rsLeader
@@ -693,6 +885,14 @@ proc serialize*(msg: RaftMessage): seq[byte] =
   stream.write(msg.leaderCommit)
   stream.write(char(if msg.success: 1 else: 0))
   stream.write(msg.matchIdx)
+  # InstallSnapshot trailing fields (appended for wire backward compatibility;
+  # pre-v1.3 peers stop reading at matchIdx and ignore these bytes)
+  stream.write(msg.snapId)
+  stream.write(msg.snapOffset)
+  stream.write(uint32(msg.snapData.len))
+  if msg.snapData.len > 0:
+    stream.writeData(addr msg.snapData[0], msg.snapData.len)
+  stream.write(char(if msg.snapDone: 1 else: 0))
   let strData = stream.data
   result = newSeq[byte](strData.len)
   for i in 0 ..< strData.len:
@@ -721,6 +921,19 @@ proc deserializeRaftMessage*(data: seq[byte]): RaftMessage =
   result.leaderCommit = stream.readUint64()
   result.success = stream.readChar() != '\0'
   result.matchIdx = stream.readUint64()
+  # Optional trailing InstallSnapshot fields (absent in pre-v1.3 buffers)
+  if not stream.atEnd:
+    result.snapId = stream.readUint64()
+  if not stream.atEnd:
+    result.snapOffset = stream.readUint64()
+  if not stream.atEnd:
+    let dataLen = int(stream.readUint32())
+    result.snapData = newSeq[byte](dataLen)
+    if dataLen > 0:
+      if stream.readData(addr result.snapData[0], dataLen) != dataLen:
+        raise newException(IOError, "Incomplete snapshot data read from stream")
+  if not stream.atEnd:
+    result.snapDone = stream.readChar() != '\0'
   stream.close()
 
 # ---------------------------------------------------------------------------
@@ -734,13 +947,16 @@ type
     running*: bool
     peerSockets*: Table[string, AsyncSocket]
     timer*: ElectionTimer
+    ## Optional TLS context; nil = plaintext (default, pre-TLS behavior).
+    tls*: TLSContext
 
-proc newRaftNetwork*(node: RaftNode): RaftNetwork =
+proc newRaftNetwork*(node: RaftNode, tls: TLSContext = nil): RaftNetwork =
   RaftNetwork(
     node: node,
     running: false,
     peerSockets: initTable[string, AsyncSocket](),
     timer: newElectionTimer(node, node.electionTimeout),
+    tls: tls,
   )
 
 const RaftConnectTimeoutMs = 200
@@ -759,6 +975,12 @@ proc connectToPeer(net: RaftNetwork, peerId: string) {.async.} =
     if not ok:
       sock.close()
       return
+    if net.tls != nil:
+      try:
+        net.tls.wrapClient(sock)
+      except CatchableError:
+        try: sock.close() except CatchableError: discard
+        return
     net.peerSockets[peerId] = sock
   except CatchableError:
     if sock != nil:
@@ -783,6 +1005,80 @@ proc broadcast*(net: RaftNetwork, msgs: seq[RaftMessage]) {.async.} =
     if i < msgs.len:
       await net.send(peer, msgs[i])
 
+proc sendSnapshot*(net: RaftNetwork, peerId: string) {.async.} =
+  ## Leader side of InstallSnapshot: build an archive of the current data dir
+  ## via the buildSnapshot callback and stream it to a lagging peer in
+  ## snapChunkBytes chunks. Triggered (via asyncCheck from processMessage)
+  ## when handleAppendReply queues the peer in snapPending after consecutive
+  ## floor-level rejects. Single-flight per peer via node.snapSending.
+  ##
+  ## Runs on the raft event loop. buildSnapshot performs the tar on the loop
+  ## under the storage gate (consistent capture); the CPU-heavy gzip then runs
+  ## on a worker thread via gzipFileAsync, awaited here, so heartbeats and the
+  ## election timer keep firing during compression instead of stalling.
+  let node = net.node
+  if peerId in node.snapSending:
+    return
+  if node.state != rsLeader or node.buildSnapshot == nil or
+      node.dataDir.len == 0:
+    return
+  let snapId = node.lastSnapshotIndex
+  if snapId == 0:
+    # snapId 0 can never be accepted (a follower's initial snapIncomingId is
+    # 0), and sends only trigger after compaction anyway — guard regardless.
+    warn("sendSnapshot: lastSnapshotIndex is 0; skipping snapshot send to " & peerId)
+    return
+  node.snapSending.incl(peerId)
+  defer: node.snapSending.excl(peerId)
+
+  let baseIndex = node.lastSnapshotIndex
+  let baseTerm = node.lastSnapshotTerm
+  # buildSnapshot writes an uncompressed tar (under the storage gate, on this
+  # loop); gzipFileAsync then compresses it on a worker thread off the loop.
+  # The follower still receives a normal .tar.gz byte stream.
+  let tarPath = node.dataDir / ("snap_out_" & $snapId & ".tar")
+  let destPath = tarPath & ".gz"
+  defer:
+    if fileExists(tarPath):
+      removeFile(tarPath)
+    if fileExists(destPath):
+      removeFile(destPath)
+
+  if not node.buildSnapshot(tarPath):
+    warn("sendSnapshot: buildSnapshot failed; aborting snapshot send to " & peerId)
+    return
+
+  if not await gzipFileAsync(tarPath, destPath):
+    warn("sendSnapshot: snapshot compression failed; aborting send to " & peerId)
+    return
+
+  var f: File
+  if not open(f, destPath, fmRead):
+    warn("sendSnapshot: cannot open built archive " & destPath)
+    return
+  defer: f.close()
+
+  let total = uint64(getFileSize(destPath))
+  var offset = 0'u64
+  while true:
+    var chunk = newSeq[byte](node.snapChunkBytes)
+    let n = f.readBytes(chunk, 0, chunk.len)
+    let done = offset + uint64(n) >= total
+    await net.send(peerId, RaftMessage(
+      kind: rmkInstallSnapshot,
+      term: node.currentTerm,
+      senderId: node.id,
+      prevLogIndex: baseIndex,   # snapshot base index/term (T7 wire layout)
+      prevLogTerm: baseTerm,
+      snapId: snapId,
+      snapOffset: offset,
+      snapData: chunk[0 ..< n],
+      snapDone: done,
+    ))
+    if done:
+      break
+    offset += uint64(n)
+
 proc processMessage*(net: RaftNetwork, msg: RaftMessage) {.async.} =
   case msg.kind
   of rmkRequestVote:
@@ -800,6 +1096,19 @@ proc processMessage*(net: RaftNetwork, msg: RaftMessage) {.async.} =
     await net.send(msg.senderId, reply)
   of rmkAppendEntriesReply:
     net.node.handleAppendReply(msg.senderId, msg)
+    # Floor-reject streak reached the threshold: this peer needs a snapshot.
+    if msg.senderId in net.node.snapPending:
+      net.node.snapPending.excl(msg.senderId)
+      asyncCheck net.sendSnapshot(msg.senderId)
+  of rmkInstallSnapshot:
+    # Same election-timer rule as AppendEntries: only a plausible current
+    # leader resets it.
+    if msg.term >= net.node.currentTerm:
+      net.timer.resetTimeout()
+    let reply = net.node.handleInstallSnapshot(msg)
+    await net.send(msg.senderId, reply)
+  of rmkInstallSnapshotReply:
+    net.node.handleInstallSnapshotReply(msg.senderId, msg)
 
 proc recvExact*(client: AsyncSocket, size: int): Future[string] {.async.} =
   ## Reads exactly `size` bytes from `client`. A short return means the peer
@@ -839,8 +1148,21 @@ proc receiveLoop(net: RaftNetwork, client: AsyncSocket) {.async.} =
 proc heartbeatLoop(net: RaftNetwork) {.async.} =
   ## Fan out heartbeats in parallel so a slow/dead peer cannot delay
   ## AppendEntries to the rest of the cluster.
+  var wasLeader = false
   while net.running:
     if net.node.state == rsLeader:
+      if not wasLeader:
+        # Fresh term, fresh connections. A peer that restarted while we were
+        # partitioned leaves a half-dead cached socket whose writes can keep
+        # "succeeding" into the void (the TCP error surfaces late or never,
+        # so the error-triggered redial in send() may never fire) — the
+        # restarted peer then never sees AppendEntries, keeps candidating,
+        # and the cluster livelocks. Drop all cached peer connections on
+        # leadership acquisition so the heartbeat fan-out redials.
+        for peerId, sock in net.peerSockets:
+          try: sock.close() except CatchableError: discard
+        net.peerSockets.clear()
+      wasLeader = true
       var futs: seq[Future[void]] = @[]
       for peer in net.node.peers:
         let msg = net.node.appendEntries(peer)
@@ -850,6 +1172,8 @@ proc heartbeatLoop(net: RaftNetwork) {.async.} =
           await f
         except CatchableError:
           discard
+    else:
+      wasLeader = false
     await sleepAsync(net.node.heartbeatTimeout)
 
 proc timerLoop*(net: RaftNetwork) {.async.}
@@ -866,6 +1190,13 @@ proc run*(net: RaftNetwork) {.async.} =
   while net.running:
     try:
       let client = await net.socket.accept()
+      if net.tls != nil:
+        try:
+          net.tls.wrapServer(client)
+        except CatchableError:
+          # Handshake failed (e.g. plaintext dial) — drop, no protocol effect.
+          client.close()
+          continue
       asyncCheck net.receiveLoop(client)
     except CatchableError:
       break

src/barabadb/core/registry.nim

@@ -203,6 +203,32 @@ proc getDatabaseInfo*(reg: DatabaseRegistry, name: string): DatabaseInfo =
     return reg.databases[name]
   return nil
 
+proc reopenDatabase*(reg: DatabaseRegistry, name: string): bool =
+  ## Reopen a database from its on-disk directory, swapping the new LSMTree
+  ## and ctx into the EXISTING DatabaseInfo slot so captured references (e.g.
+  ## the raft applyCommand closure) see the new state.
+  ## Minimal API added for raft InstallSnapshot restore: the caller must have
+  ## closed info.db first (snapshot restore closes it before swapping the data
+  ## directory); this proc does not close.
+  ## Returns false if the database is unknown or the reopen fails.
+  acquire(reg.lock)
+  let info = if name in reg.databases: reg.databases[name] else: nil
+  release(reg.lock)
+  if info == nil:
+    return false
+  try:
+    let dbDir = reg.dataRoot / name
+    let db = openLsmForRegistry(reg, dbDir)
+    let ctx = reg.ctxFactory(db, reg)
+    info.db = db
+    info.ctx = ctx
+    return true
+  except CatchableError as e:
+    # echo instead of logging: callers include gcsafe raft callbacks, and
+    # core/logging's info/warn are not gcsafe.
+    echo "[registry] Error reopening database '", name, "': ", e.msg
+    return false
+
 proc closeAll*(reg: DatabaseRegistry) =
   acquire(reg.lock)
   defer: release(reg.lock)

src/barabadb/core/replication.nim

@@ -96,6 +96,41 @@ proc connectWithTimeout(sock: Socket, host: string, port: Port, timeoutMs: int):
     sock.getFd.setBlocking(true)
     return err == 0
 
+type
+  RepOp* = enum
+    ropPut
+    ropDelete
+    ropInvalid
+
+proc encodeRepPayload*(deleted: bool, key: string, value: seq[byte]): seq[byte] =
+  ## Tagged legacy-REP payload. A leading op tag makes put/delete explicit so
+  ## an empty put value (PK-only rows store an empty LSM value) is never
+  ## mistaken for a delete on the receiver:
+  ##   put    -> 'P' & key & "\x00" & value   (value may be empty)
+  ##   delete -> 'D' & key
+  ## Mirrors the raft convention (explicit "put"/"delete" commands).
+  if deleted:
+    cast[seq[byte]]("D" & key)
+  else:
+    cast[seq[byte]]("P" & key & "\x00" & cast[string](value))
+
+proc decodeRepPayload*(data: seq[byte]): tuple[op: RepOp, key: string, value: seq[byte]] =
+  ## Inverse of encodeRepPayload. Returns ropInvalid for empty or untagged
+  ## payloads rather than guessing the operation from the value length.
+  if data.len == 0:
+    return (ropInvalid, "", @[])
+  case char(data[0])
+  of 'P':
+    let body = data[1 ..< data.len]
+    let nullPos = find(body, byte(0))
+    if nullPos < 0:
+      return (ropInvalid, "", @[])
+    return (ropPut, cast[string](body[0 ..< nullPos]), body[nullPos + 1 ..< body.len])
+  of 'D':
+    return (ropDelete, cast[string](data[1 ..< data.len]), @[])
+  else:
+    return (ropInvalid, "", @[])
+
 proc shipToReplica(replica: Replica, lsn: uint64, data: seq[byte]): bool =
   ## Send replication data to a replica via TCP.
   ## Protocol: "REP <lsn> <dataLen>\n<data>"
@@ -178,10 +213,18 @@ proc writeLsn*(rm: ReplicationManager, data: seq[byte]): uint64 =
         rm.pendingAcks[lsn].excl(id)
       if rm.pendingAcks[lsn].len == 0:
         rm.pendingAcks.del(lsn)
+    # Semi-sync requires at least syncReplicaCount acks when replicas are
+    # connected. With zero connected peers (nothing to ship) the write is
+    # local-only — same as sync mode with an empty replica set.
+    if rm.syncReplicaCount > 0 and replicasToShip.len > 0 and
+       ackCount < rm.syncReplicaCount:
+      # Drop the LSN from pendingAcks — write is not durable
+      rm.pendingAcks.del(lsn)
+      release(rm.lock)
+      echo "[ERROR] Semi-sync replication failed: only ", ackCount, "/",
+           rm.syncReplicaCount, " replicas acked for LSN ", lsn
+      return 0
     release(rm.lock)
-    if replicasToShip.len > 0 and ackCount == 0 and rm.syncReplicaCount > 0:
-      when defined(debug):
-        echo "Replication semi-sync: no replicas acked for LSN ", lsn
     return lsn
 
 proc ackLsn*(rm: ReplicationManager, replicaId: string, lsn: uint64) =

src/barabadb/core/server.nim

@@ -1,6 +1,7 @@
 ## BaraDB Server — async TCP server with wire protocol
 import std/asyncdispatch
 import std/asyncnet
+import std/os
 import std/strutils
 import std/sequtils
 import std/tables
@@ -22,6 +23,7 @@ import ../query/parser
 import ../query/ast
 import ../query/executor
 import ../query/exec/params
+import ../query/exec/dml
 import ../storage/lsm
 import ../storage/gate
 import ../core/mvcc
@@ -49,6 +51,10 @@ type
     clusterMembership*: ClusterMembership
     gossipProtocol*: GossipProtocol
     tls*: TLSContext
+    ## Dedicated client-role TLS context for follower→leader forwarding.
+    ## Must not reuse `tls` with verifyPeer — OpenSSL contexts are role-agnostic
+    ## and enabling verify on the server context would break inbound handshakes.
+    tlsClient*: TLSContext
     rateLimiter*: RateLimiter
     activeConnections*: int
     activeConnectionsLock*: Lock
@@ -65,9 +71,22 @@ proc newServerWithRegistry*(config: BaraConfig, registry: DatabaseRegistry): Ser
   let ctx = cast[ExecutionContext](cast[pointer](dbInfo.ctx))
   ctx.txnManager = newTxnManager()
   var tls: TLSContext = nil
+  var tlsClient: TLSContext = nil
   if config.tlsEnabled and config.certFile.len > 0 and config.keyFile.len > 0:
-    let tlsConfig = newTLSConfig(config.certFile, config.keyFile)
-    tls = newTLSContext(tlsConfig)
+    if config.tlsVerifyPeer and config.tlsCaFile.len == 0:
+      raise newException(ValueError,
+        "tlsVerifyPeer is true but CA file is missing. Set BARADB_TLS_CA_FILE")
+    if config.tlsVerifyPeer and config.tlsCaFile.len > 0 and
+       not fileExists(config.tlsCaFile):
+      raise newException(ValueError,
+        "BARADB_TLS_VERIFY_PEER=true but CA file missing: " & config.tlsCaFile)
+    tls = newTLSContext(newTLSConfig(config.certFile, config.keyFile))
+    if config.tlsVerifyPeer:
+      tlsClient = newTLSContext(newTLSConfig(
+        config.certFile, config.keyFile,
+        caFile = config.tlsCaFile, verifyPeer = true))
+    else:
+      tlsClient = tls
 
   # Initialize sharding / gossip. Server fields own the refs; locals used inside
   # callback closures are {.cursor.} so ARC does not form uncollectable cycles
@@ -84,6 +103,7 @@ proc newServerWithRegistry*(config: BaraConfig, registry: DatabaseRegistry): Ser
          clusterMembership: nil,
          gossipProtocol: newGossipProtocol(localId, config.address, config.port, gossipPort = gossipPort),
          tls: tls,
+         tlsClient: tlsClient,
          rateLimiter: rl)
   result.clusterMembership = newClusterMembership(result.shardRouter, localId)
   initLock(result.activeConnectionsLock)
@@ -165,6 +185,11 @@ proc parseHeader(data: string): (bool, MessageHeader) =
     return (false, MessageHeader())
   let kind = cast[MsgKind](rawKind)
   let length = readUint32BE(data, 4)
+  # Reject oversized messages before any buffer allocation: recvExactWithTimeout
+  # pre-allocates `length` bytes before the auth check, so an unbounded uint32
+  # (up to ~4 GiB) is a pre-auth memory-exhaustion DoS. Cap at the wire max.
+  if length > uint32(MaxWireStringLen):
+    return (false, MessageHeader())
   let requestId = readUint32BE(data, 8)
   return (true, MessageHeader(kind: kind, length: length, requestId: requestId))
 
@@ -216,16 +241,28 @@ proc forwardRecvExact(sock: AsyncSocket, size: int): Future[string] {.async.} =
   return buf
 
 proc forwardQueryToLeader*(host: string, port: int, query: string,
+                           tls: TLSContext = nil,
                            params: seq[WireValue] = @[],
                            timeoutMs: int = 5000): Future[(bool, QueryResult, string)] {.async.} =
   ## Proxy a write/DDL to the known leader's SQL port. Used by followers when
   ## BARADB_RAFT_CLIENT_PEERS maps leader id → host:clientPort.
+  ## `tls` is a *client-role* context (Server.tlsClient). When the wire port
+  ## serves TLS, the leader's does too, so the forwarding dial must complete a
+  ## client handshake. Peer verification is honoured when that context was
+  ## built with verifyPeer (BARADB_TLS_VERIFY_PEER + BARADB_TLS_CA_FILE).
+  ## Do NOT pass Server.tls (the inbound/server context) with verifyPeer
+  ## flipped on — OpenSSL contexts are role-agnostic in Nim's stdlib.
   var sock: AsyncSocket = nil
   try:
     sock = newAsyncSocket()
     let okConn = await withTimeout(sock.connect(host, Port(port)), min(timeoutMs, 2000))
     if not okConn:
       return (false, QueryResult(), "leader forward connect timeout")
+    if tls != nil:
+      try:
+        tls.wrapClient(sock)
+      except CatchableError:
+        return (false, QueryResult(), "leader forward TLS handshake failed")
     let reqId = 1'u32
     let msg = if params.len > 0:
         makeQueryParamsMessage(reqId, query, params)
@@ -306,10 +343,12 @@ proc waitRaftCommit(node: RaftNode, lastIdx: uint64, timeoutMs: int): Future[(bo
     node.metrics.commitWaitMsTotal += waitedMs
   return (true, "")
 
-proc appendWriteToRaft*(node: RaftNode, kvPairs: seq[(string, seq[byte])],
+proc appendWriteToRaft*(node: RaftNode,
+                        kvPairs: seq[tuple[key: string, value: seq[byte], deleted: bool]],
                         timeoutMs: int): Future[(bool, string)] {.async.} =
   ## C3b leader write path: append each written KV pair to the Raft log and
-  ## wait for majority commit. An empty value encodes a delete; the entry
+  ## wait for majority commit. The `deleted` flag encodes a delete — an empty
+  ## value alone is a put (PK-only tables store an empty LSM value); the entry
   ## format matches applyCommand ("put": key \x00 value, "delete": key).
   ##
   ## MUST be called from the async event-loop thread that owns `node` and
@@ -317,11 +356,11 @@ proc appendWriteToRaft*(node: RaftNode, kvPairs: seq[(string, seq[byte])],
   ## handleAppendReply on the same loop, and applyCommand re-enters the
   ## (non-reentrant) gate — waiting under the gate would deadlock the loop.
   var lastIdx = 0'u64
-  for (key, value) in kvPairs:
-    let entry = if value.len > 0:
-        node.appendLog("put", cast[seq[byte]](key & "\x00" & cast[string](value)))
+  for pair in kvPairs:
+    let entry = if pair.deleted:
+        node.appendLog("delete", cast[seq[byte]](pair.key))
       else:
-        node.appendLog("delete", cast[seq[byte]](key))
+        node.appendLog("put", cast[seq[byte]](pair.key & "\x00" & cast[string](pair.value)))
     if entry.index == 0:
       if node.metrics != nil:
         inc node.metrics.lostLeadershipTotal
@@ -346,14 +385,15 @@ proc executeQuery(db: LSMTree, ctx: ExecutionContext, query: string, params: seq
                    raftNode: RaftNode = nil,
                    raftWriteTimeoutMs: int = 5000,
                    raftPeerClientAddrs: Table[string, tuple[host: string, port: int]] =
-                     initTable[string, tuple[host: string, port: int]]()): Future[(bool, QueryResult, string)] {.async.} =
+                     initTable[string, tuple[host: string, port: int]](),
+                   forwardTls: TLSContext = nil): Future[(bool, QueryResult, string)] {.async.} =
   ## All storage access is under the global StorageGate so HTTP worker threads
   ## and the TCP event loop never touch ORC-managed LSM/executor state concurrently.
   ## The gate is released BEFORE the Raft commit wait — see appendWriteToRaft.
   var ok = false
   var qr = QueryResult()
   var msg = ""
-  var kvPairs: seq[(string, seq[byte])] = @[]
+  var kvPairs: seq[tuple[key: string, value: seq[byte], deleted: bool]] = @[]
   var needsRaftDdl = false
   var needsForward = false
   var forwardHost = ""
@@ -397,12 +437,13 @@ proc executeQuery(db: LSMTree, ctx: ExecutionContext, query: string, params: seq
           # Ship written key-value pairs to replicas (legacy path; skipped when
           # the raft path below handles the statement).
           if raftNode == nil and replication != nil and res.keyValuePairs.len > 0:
-            for (key, value) in res.keyValuePairs:
-              var data = newSeq[byte](key.len + 1 + value.len)
-              for i, c in key: data[i] = byte(c)
-              data[key.len] = byte(0)
-              for i, c in value: data[key.len + 1 + i] = c
-              discard replication.writeLsn(data)
+            for pair in res.keyValuePairs:
+              # Legacy REP wire format: explicit 'P'/'D' op tag (see
+              # encodeRepPayload). The tag — not an empty value — distinguishes
+              # a put from a delete, so PK-only rows (empty value) replicate as
+              # puts instead of vanishing as deletes.
+              discard replication.writeLsn(
+                encodeRepPayload(pair.deleted, pair.key, pair.value))
           qr = QueryResult(affectedRows: res.affectedRows, rowCount: res.rows.len)
           qr.columns = res.columns
 
@@ -446,7 +487,7 @@ proc executeQuery(db: LSMTree, ctx: ExecutionContext, query: string, params: seq
   # Follower write/DDL: proxy to leader SQL port (outside the storage gate).
   if needsForward:
     let (okF, qrF, errF) = await forwardQueryToLeader(forwardHost, forwardPort,
-      query, params, raftWriteTimeoutMs)
+      query, forwardTls, params, raftWriteTimeoutMs)
     if raftNode != nil and raftNode.metrics != nil:
       if okF: inc raftNode.metrics.forwardsTotal
       else: inc raftNode.metrics.forwardErrorsTotal
@@ -637,14 +678,25 @@ proc handleClient(server: Server, client: AsyncSocket, clientId: int) {.async.}
               if chunk.len == 0: break
               data.add(chunk)
             if data.len > 0:
-              let nullPos = data.find('\0')
-              if nullPos >= 0:
-                let key = data[0..<nullPos]
-                let value = data[nullPos+1..^1]
-                if value.len > 0:
-                  server.db.put(key, stringToBytes(value))
-                else:
-                  server.db.delete(key)
+              # Op tag — not value length — decides put vs delete, so a PK-only
+              # put (empty value) is applied as a put and the row survives.
+              let decoded = decodeRepPayload(cast[seq[byte]](data))
+              case decoded.op
+              of ropPut, ropDelete:
+                # Apply through applyReplicatedPut/Delete (not raw db.put/delete)
+                # so secondary B-tree/FTS/HNSW/graph indexes stay consistent on
+                # the replica — the same path raft uses. server.ctx is the
+                # canonical default ctx whose index structures the per-connection
+                # query clones share. Under the storage gate: those structures
+                # are shared with hunos HTTP workers and are only safe to mutate
+                # under it.
+                withStorageGate:
+                  if decoded.op == ropPut:
+                    applyReplicatedPut(server.ctx, decoded.key, decoded.value)
+                  else:
+                    applyReplicatedDelete(server.ctx, decoded.key)
+              of ropInvalid:
+                discard
           await client.send("ACK " & $lsn & "\n")
         else:
           await client.send("ERR\n")
@@ -652,6 +704,9 @@ proc handleClient(server: Server, client: AsyncSocket, clientId: int) {.async.}
 
       # Detect shard migration data (starts with "MIGRATE ")
       if headerData.len >= 8 and headerData[0..7] == "MIGRATE ":
+        if not authenticated:
+          await client.send("ERR auth required\n")
+          continue
         var rest = headerData[8..^1]
         while '\n' notin rest:
           let more = await client.recvWithTimeout(1024, idleTimeout)
@@ -769,7 +824,8 @@ proc handleClient(server: Server, client: AsyncSocket, clientId: int) {.async.}
           let (success, result, errorMsg) = await executeQuery(connCtx.db, connCtx, queryStr,
             replication=server.replicationManager, raftNode=server.raftNode,
             raftWriteTimeoutMs=server.config.raftWriteTimeoutMs,
-            raftPeerClientAddrs=server.config.raftPeerClientAddrs)
+            raftPeerClientAddrs=server.config.raftPeerClientAddrs,
+            forwardTls=server.tlsClient)
           let durationMs = int((getMonoTime().ticks() - startTicks) div 1_000_000)
 
           if durationMs >= slowThreshold:
@@ -792,7 +848,8 @@ proc handleClient(server: Server, client: AsyncSocket, clientId: int) {.async.}
         let (success, result, errorMsg) = await executeQuery(connCtx.db, connCtx, queryStr, params,
           replication=server.replicationManager, raftNode=server.raftNode,
           raftWriteTimeoutMs=server.config.raftWriteTimeoutMs,
-          raftPeerClientAddrs=server.config.raftPeerClientAddrs)
+          raftPeerClientAddrs=server.config.raftPeerClientAddrs,
+          forwardTls=server.tlsClient)
         let durationMs = int((getMonoTime().ticks() - startTicks) div 1_000_000)
 
         if durationMs >= slowThreshold:

src/barabadb/core/websocket.nim

@@ -15,18 +15,25 @@ else:
 import config
 import jwt as jwtlib
 
+const
+  ## RFC 6455: reject oversized frames/messages to bound memory.
+  MaxWsFrameBytes* = 1 * 1024 * 1024
+  MaxWsMessageBytes* = 4 * 1024 * 1024
+  MaxWsControlPayload* = 125
+
 type
-  WsFrame = object
-    fin: bool
-    opcode: uint8
-    masked: bool
-    payloadLen: uint64
-    maskKey: array[4, byte]
-    payload: string
+  WsFrame* = object
+    fin*: bool
+    opcode*: uint8
+    masked*: bool
+    payloadLen*: uint64
+    maskKey*: array[4, byte]
+    payload*: string
 
   WsClient* = ref object
     socket: AsyncSocket
     id: int
+    username: string
     subscriptions: HashSet[string]
 
   WsServer* = ref object
@@ -35,6 +42,8 @@ type
     running: bool
     config*: BaraConfig
     secretKey*: string
+    ## Table-level read authorization for SUBSCRIBE. Nil + authEnabled → deny.
+    canSubscribe*: proc(username, table: string): bool {.closure, gcsafe.}
     onInsert*: proc (table, key, value: string) {.closure.}
     onDelete*: proc (table, key: string) {.closure.}
 
@@ -46,20 +55,19 @@ proc newWsServer*(cfg: BaraConfig = defaultConfig(), secret: string = ""): WsSer
 # WebSocket frame encoding/decoding (RFC 6455)
 # ----------------------------------------------------------------------
 
-proc encodeFrame(opcode: uint8, payload: string): string =
+proc encodeFrame*(opcode: uint8, payload: string, masked = false,
+                  maskKey: array[4, byte] = [0'u8, 0, 0, 0]): string =
   result = ""
-  let isMasked = false
   var b0 = 0x80'u8 or opcode
   result.add(char(b0))
 
-  var b1 = 0'u8
-  if not isMasked:
-    if payload.len < 126:
-      b1 = uint8(payload.len)
-    elif payload.len <= 65535:
-      b1 = 126
-    else:
-      b1 = 127
+  var b1 = if masked: 0x80'u8 else: 0'u8
+  if payload.len < 126:
+    b1 = b1 or uint8(payload.len)
+  elif payload.len <= 65535:
+    b1 = b1 or 126
+  else:
+    b1 = b1 or 127
   result.add(char(b1))
 
   if payload.len >= 126 and payload.len <= 65535:
@@ -71,9 +79,17 @@ proc encodeFrame(opcode: uint8, payload: string): string =
     for i in countdown(7, 0):
       result.add(char((len64 shr (i * 8)) and 0xFF))
 
-  result.add(payload)
+  if masked:
+    for i in 0..3:
+      result.add(char(maskKey[i]))
+    for i, c in payload:
+      result.add(char(byte(c) xor maskKey[i mod 4]))
+  else:
+    result.add(payload)
 
-proc decodeFrame(data: string): (WsFrame, int) =
+proc decodeFrame*(data: string): (WsFrame, int) =
+  ## Returns (frame, consumed). consumed == 0 → need more bytes;
+  ## consumed < 0 → protocol error (close the connection).
   if data.len < 2:
     return (WsFrame(), 0)
 
@@ -84,6 +100,10 @@ proc decodeFrame(data: string): (WsFrame, int) =
   frame.opcode = b0 and 0x0F
   frame.masked = (b1 and 0x80) != 0
 
+  # RFC 6455 §5.1 — client-to-server frames MUST be masked.
+  if not frame.masked:
+    return (WsFrame(), -1)
+
   var len = uint64(b1 and 0x7F)
   var offset = 2
 
@@ -98,23 +118,24 @@ proc decodeFrame(data: string): (WsFrame, int) =
       len = (len shl 8) or uint64(uint8(data[2 + i]))
     offset = 10
 
-  if frame.masked:
-    if data.len < offset + 4: return (WsFrame(), 0)
-    for i in 0..3:
-      frame.maskKey[i] = byte(data[offset + i])
-    offset += 4
+  let isControl = frame.opcode == 0x8 or frame.opcode == 0x9 or frame.opcode == 0xA
+  if isControl and (not frame.fin or len > uint64(MaxWsControlPayload)):
+    return (WsFrame(), -1)
+  if len > uint64(MaxWsFrameBytes):
+    return (WsFrame(), -1)
+
+  if data.len < offset + 4: return (WsFrame(), 0)
+  for i in 0..3:
+    frame.maskKey[i] = byte(data[offset + i])
+  offset += 4
 
   if uint64(data.len) < uint64(offset) + len:
-    return (Wsframe(), 0)
+    return (WsFrame(), 0)
 
-  if len > uint64(high(int) - 1):
-    return (Wsframe(), 0)
   let plen = int(len)
-  if frame.masked:
-    for i in 0..<plen:
-      frame.payload.add(char(byte(data[offset + i]) xor frame.maskKey[i mod 4]))
-  else:
-    frame.payload = data[offset..offset + plen - 1]
+  frame.payloadLen = len
+  for i in 0..<plen:
+    frame.payload.add(char(byte(data[offset + i]) xor frame.maskKey[i mod 4]))
 
   return (frame, offset + plen)
 
@@ -179,6 +200,14 @@ proc computeAcceptKey(key: string): string =
 # Subscription management
 # ----------------------------------------------------------------------
 
+proc validSubscribeTable(table: string): bool =
+  if table.len == 0 or table.len > 128:
+    return false
+  for c in table:
+    if c notin {'a'..'z', 'A'..'Z', '0'..'9', '_'}:
+      return false
+  return true
+
 proc subscribe*(client: WsClient, table: string) =
   client.subscriptions.incl(table)
 
@@ -201,9 +230,11 @@ proc broadcastToTable*(server: WsServer, table: string, msg: string) {.async.} =
 # WebSocket client handler
 # ----------------------------------------------------------------------
 
-proc handleWsClient(server: WsServer, client: AsyncSocket, id: int) {.async.} =
+proc handleWsClient(server: WsServer, client: AsyncSocket, id: int,
+                    username: string = "") {.async.} =
   echo "WebSocket client ", id, " connected"
-  var wsClient = WsClient(socket: client, id: id, subscriptions: initHashSet[string]())
+  var wsClient = WsClient(socket: client, id: id, username: username,
+                          subscriptions: initHashSet[string]())
   server.clients[id] = wsClient
 
   var buf = ""
@@ -212,12 +243,22 @@ proc handleWsClient(server: WsServer, client: AsyncSocket, id: int) {.async.} =
       let chunk = await client.recv(4096)
       if chunk.len == 0:
         break
+      if buf.len + chunk.len > MaxWsMessageBytes:
+        let closeF = encodeFrame(0x8, "")
+        try: await client.send(closeF) except CatchableError: discard
+        break
       buf.add(chunk)
 
       while buf.len >= 2:
         let (frame, consumed) = decodeFrame(buf)
         if consumed == 0:
           break
+        if consumed < 0:
+          let closeF = encodeFrame(0x8, "")
+          try: await client.send(closeF) except CatchableError: discard
+          client.close()
+          server.clients.del(id)
+          return
 
         case frame.opcode
         of 0x8:  # close
@@ -231,9 +272,18 @@ proc handleWsClient(server: WsServer, client: AsyncSocket, id: int) {.async.} =
           let msg = frame.payload
           if msg.startsWith("SUBSCRIBE "):
             let table = msg[10..^1].strip()
-            wsClient.subscribe(table)
-            let ack = encodeFrame(0x1, "OK subscribed to " & table)
-            await client.send(ack)
+            var allowed = validSubscribeTable(table)
+            if allowed and server.config.authEnabled:
+              if username.len == 0 or server.canSubscribe == nil or
+                 not server.canSubscribe(username, table):
+                allowed = false
+            if allowed:
+              wsClient.subscribe(table)
+              let ack = encodeFrame(0x1, "OK subscribed to " & table)
+              await client.send(ack)
+            else:
+              let nack = encodeFrame(0x1, "ERR subscribe denied for " & table)
+              await client.send(nack)
           elif msg.startsWith("UNSUBSCRIBE "):
             let table = msg[12..^1].strip()
             wsClient.unsubscribe(table)
@@ -285,6 +335,7 @@ proc handleConnection(server: WsServer, client: AsyncSocket) {.async.} =
     return
 
   # Auth check
+  var username = ""
   if server.config.authEnabled:
     let authHeader = headers.getOrDefault("authorization", "")
     if authHeader.len == 0 or not authHeader.startsWith("Bearer "):
@@ -305,6 +356,8 @@ proc handleConnection(server: WsServer, client: AsyncSocket) {.async.} =
           await client.send("HTTP/1.1 401 Unauthorized\r\n\r\n")
           client.close()
           return
+      if "sub" in token.claims:
+        username = token.claims["sub"].node.str
     except CatchableError:
       await client.send("HTTP/1.1 401 Unauthorized\r\n\r\n")
       client.close()
@@ -321,7 +374,7 @@ proc handleConnection(server: WsServer, client: AsyncSocket) {.async.} =
   await client.send(response)
 
   inc server.nextId
-  asyncCheck server.handleWsClient(client, server.nextId)
+  asyncCheck server.handleWsClient(client, server.nextId, username)
 
 proc setTcpNoDelay(sock: AsyncSocket) =
   ## Enable TCP_NODELAY using the correct protocol level (IPPROTO_TCP).

src/barabadb/protocol/auth.nim

@@ -221,10 +221,17 @@ proc registerScramUser*(am: AuthManager, username, password: string,
   let cred = createScramCredential(password, iterationCount = iterationCount)
   am.scramUsers[username] = cred
 
+proc dummyScramStartWork() =
+  ## Match known-user startScram work (urandom nonce + base64) so unknown
+  ## users cannot be enumerated by timing.
+  discard generateNonce()
+  discard encode("0123456789abcdef0123456789abcdef")
+
 proc startScram*(am: AuthManager, clientFirstMessage: string): string =
   ## Start SCRAM authentication. Returns server-first-message.
-  let (_, username, clientNonce) = parseClientFirst(clientFirstMessage)
-  if username notin am.scramUsers:
+  let (gs2, username, clientNonce) = parseClientFirst(clientFirstMessage)
+  if username notin am.scramUsers or gs2 notin ["n", "y"]:
+    dummyScramStartWork()
     raise newException(ValueError, "Authentication failed")
 
   let cred = am.scramUsers[username]
@@ -239,6 +246,7 @@ proc startScram*(am: AuthManager, clientFirstMessage: string): string =
 
   var state = ScramServerState(
     username: username,
+    gs2Flag: gs2,
     clientFirstMessageBare: clientFirstMessageBare,
     serverFirstMessage: serverFirst,
     authMessage: authMessage,
@@ -264,6 +272,9 @@ proc finishScram*(am: AuthManager, clientFinalMessage: string): (bool, string) =
   var state = am.scramSessions[nonce]
   am.scramSessions.del(nonce)
 
+  if stripB64Padding(cbind) != expectedChannelBinding(state.gs2Flag):
+    return (false, "e=channel-bindings-dont-match")
+
   # Update authMessage with client-final-message-without-proof
   let clientFinalWithoutProof = "c=" & cbind & ",r=" & nonce
   state.authMessage = state.authMessage & "," & clientFinalWithoutProof

src/barabadb/protocol/scram.nim

@@ -19,6 +19,7 @@ type
 
   ScramServerState* = object
     username*: string
+    gs2Flag*: string
     clientFirstMessageBare*: string
     serverFirstMessage*: string
     authMessage*: string
@@ -189,6 +190,19 @@ proc createScramCredential*(password: string, salt: string = "",
 # SCRAM message parsing / building
 # ---------------------------------------------------------------------------
 
+proc expectedChannelBinding*(gs2Flag: string): string =
+  ## RFC 5802 cbind-input with no cbind-data is the gs2-header: flag + "," + authzid + ",".
+  ## Authzid is unsupported, so the header is `n,,` or `y,,`.
+  let header = gs2Flag & ",,"
+  result = encode(header)
+  while result.endsWith("="):
+    result.setLen(result.len - 1)
+
+proc stripB64Padding*(s: string): string =
+  result = s
+  while result.endsWith("="):
+    result.setLen(result.len - 1)
+
 proc parseClientFirst*(msg: string): (string, string, string) =
   ## Parse client-first-message: gs2-header,username,nonce
   ## Returns: (gs2_header, username, nonce)

src/barabadb/protocol/ssl.nim

@@ -29,10 +29,14 @@ proc newTLSConfig*(certFile: string, keyFile: string, caFile: string = "",
 proc newTLSContext*(config: TLSConfig): TLSContext =
   result = TLSContext(config: config)
   if fileExists(config.certFile) and fileExists(config.keyFile):
+    # caFile is only honored by newContext when verifyPeer is true
+    # (verifyMode != CVerifyNone); a missing CA file then raises IOError,
+    # which is the desired fail-closed behavior.
     result.sslCtx = newContext(
       certFile = config.certFile,
       keyFile = config.keyFile,
       verifyMode = if config.verifyPeer: CVerifyPeer else: CVerifyNone,
+      caFile = config.caFile,
     )
   else:
     raise newException(IOError, "TLS certificate or key file not found: " &
@@ -40,7 +44,11 @@ proc newTLSContext*(config: TLSConfig): TLSContext =
 
 proc wrapClient*(tls: TLSContext, socket: AsyncSocket) {.inline.} =
   if tls.sslCtx != nil:
-    tls.sslCtx.wrapSocket(socket)
+    # wrapConnectedSocket (asyncnet overload) sets connect state; the
+    # handshake itself is driven lazily by the first send/recv. Plain
+    # wrapSocket leaves the SSL handle in SSL_ST_BEFORE and the first
+    # SSL_write fails with "uninitialized".
+    tls.sslCtx.wrapConnectedSocket(socket, handshakeAsClient)
 
 proc wrapServer*(tls: TLSContext, socket: AsyncSocket) {.inline.} =
   if tls.sslCtx != nil:

src/barabadb/query/exec/dml.nim

@@ -45,7 +45,7 @@ proc violatesUniqueIndex*(ctx: ExecutionContext, table: string, fields: seq[stri
   return ""
 
 proc execInsert*(ctx: ExecutionContext, table: string, fields: seq[string], values: seq[seq[string]],
-                  kvPairs: var seq[(string, seq[byte])]): int =
+                  kvPairs: var seq[tuple[key: string, value: seq[byte], deleted: bool]]): int =
   if not hasPrivilege(ctx, table, "INSERT"):
     return 0
   let tblDef = if table in ctx.tables: ctx.tables[table] else: TableDef()
@@ -87,7 +87,7 @@ proc execInsert*(ctx: ExecutionContext, table: string, fields: seq[string], valu
       discard ctx.txnManager.write(ctx.pendingTxn, fullKey, cast[seq[byte]](valStr))
     else:
       ctx.db.put(fullKey, cast[seq[byte]](valStr))
-      kvPairs.add((fullKey, cast[seq[byte]](valStr)))
+      kvPairs.add((fullKey, cast[seq[byte]](valStr), false))
 
     for colName in ctx.btrees.keys.toSeq():
       if colName.startsWith(table & "."):
@@ -221,7 +221,7 @@ proc execInsert*(ctx: ExecutionContext, table: string, fields: seq[string], valu
   return count
 
 proc execDelete*(ctx: ExecutionContext, table: string, key: string,
-                  kvPairs: var seq[(string, seq[byte])]): int =
+                  kvPairs: var seq[tuple[key: string, value: seq[byte], deleted: bool]]): int =
   if not hasPrivilege(ctx, table, "DELETE"):
     return 0
   let fullKey = table & "." & key
@@ -238,7 +238,7 @@ proc execDelete*(ctx: ExecutionContext, table: string, key: string,
       discard ctx.txnManager.delete(ctx.pendingTxn, fullKey)
     else:
       ctx.db.delete(fullKey)
-      kvPairs.add((fullKey, @[]))
+      kvPairs.add((fullKey, @[], true))
     # Update BTree indexes
     for colName in ctx.btrees.keys.toSeq():
       if colName.startsWith(table & "."):
@@ -264,7 +264,7 @@ proc execDelete*(ctx: ExecutionContext, table: string, key: string,
   return 0
 
 proc execUpdateRow*(ctx: ExecutionContext, table: string, key: string, sets: Table[string, string],
-                     kvPairs: var seq[(string, seq[byte])]): int =
+                     kvPairs: var seq[tuple[key: string, value: seq[byte], deleted: bool]]): int =
   if not hasPrivilege(ctx, table, "UPDATE"):
     return 0
   let fullKey = table & "." & key
@@ -313,7 +313,7 @@ proc execUpdateRow*(ctx: ExecutionContext, table: string, key: string, sets: Tab
     discard ctx.txnManager.write(ctx.pendingTxn, fullKey, cast[seq[byte]](newVal))
   else:
     ctx.db.put(fullKey, cast[seq[byte]](newVal))
-    kvPairs.add((fullKey, cast[seq[byte]](newVal)))
+    kvPairs.add((fullKey, cast[seq[byte]](newVal), false))
   # Update FTS indexes: remove old doc, add new
   for ftsKey, ftsIdx in ctx.ftsIndexes:
     if ftsKey.startsWith(table & "."):

src/barabadb/query/exec/eval.nim

@@ -429,6 +429,8 @@ proc evalExprOld*(expr: IRExpr, row: Table[string, string], ctx: ExecutionContex
     let right = evalExprOld(expr.binRight, row, ctx)
     case expr.binOp
     of irEq:
+      # SQL three-valued logic: any NULL operand → unknown, not true.
+      if isNull(left) or isNull(right): return "\\N"
       if left == right: return "true"
       # Try numeric comparison
       try:
@@ -436,32 +438,45 @@ proc evalExprOld*(expr: IRExpr, row: Table[string, string], ctx: ExecutionContex
       except CatchableError: discard
       return "false"
     of irNeq:
-      if left != right: return "true"
-      # Try numeric comparison
+      if isNull(left) or isNull(right): return "\\N"
+      # Numeric-first so `!=` is the exact complement of `=` (irEq): string
+      # inequality alone would make `5 != 5.0` true while `5 = 5.0` is true.
       try:
         return if parseFloat(left) != parseFloat(right): "true" else: "false"
-      except CatchableError: return "false"
+      except CatchableError:
+        return if left != right: "true" else: "false"
     of irLt:
+      if isNull(left) or isNull(right): return "\\N"
       try:
         return if parseFloat(left) < parseFloat(right): "true" else: "false"
       except CatchableError: return if left < right: "true" else: "false"
     of irLte:
+      if isNull(left) or isNull(right): return "\\N"
       try:
         return if parseFloat(left) <= parseFloat(right): "true" else: "false"
       except CatchableError: return if left <= right: "true" else: "false"
     of irGt:
+      if isNull(left) or isNull(right): return "\\N"
       try:
         return if parseFloat(left) > parseFloat(right): "true" else: "false"
       except CatchableError: return if left > right: "true" else: "false"
     of irGte:
+      if isNull(left) or isNull(right): return "\\N"
       try:
         return if parseFloat(left) >= parseFloat(right): "true" else: "false"
       except CatchableError: return if left >= right: "true" else: "false"
     of irAnd:
-      if left == "true" and right == "true": return "true"
-      return "false"
+      # false AND x = false; unknown AND true/unknown = unknown; else both true.
+      let lNull = isNull(left)
+      let rNull = isNull(right)
+      let lTrue = left == "true"
+      let rTrue = right == "true"
+      if (not lNull and not lTrue) or (not rNull and not rTrue): return "false"
+      if lNull or rNull: return "\\N"
+      return "true"
     of irOr:
       if left == "true" or right == "true": return "true"
+      if isNull(left) or isNull(right): return "\\N"
       return "false"
     of irAdd, irSub, irMul, irDiv, irMod, irPow:
       let v = evalExpr(expr, stringTableToValueRow(row), ctx)
@@ -474,6 +489,7 @@ proc evalExprOld*(expr: IRExpr, row: Table[string, string], ctx: ExecutionContex
       of vkString: return v.strVal
       else: return "\\N"
     of irLike:
+      if isNull(left) or isNull(right): return "\\N"
       proc escapeRe(s: string): string =
         result = ""
         for ch in s:
@@ -489,6 +505,7 @@ proc evalExprOld*(expr: IRExpr, row: Table[string, string], ctx: ExecutionContex
       except CatchableError: discard
       return "false"
     of irILike:
+      if isNull(left) or isNull(right): return "\\N"
       proc escapeRe(s: string): string =
         result = ""
         for ch in s:
@@ -504,8 +521,10 @@ proc evalExprOld*(expr: IRExpr, row: Table[string, string], ctx: ExecutionContex
       except CatchableError: discard
       return "false"
     of irIn:
+      if isNull(left): return "\\N"
       if expr.binRight.kind == irekSubquery:
         let subRows = requireExecutePlanHook()(ctx, expr.binRight.subqueryPlan)
+        var sawNull = false
         for row in subRows:
           # Compare against the first non-internal column only (SQL semantics)
           var firstVal = ""
@@ -515,8 +534,14 @@ proc evalExprOld*(expr: IRExpr, row: Table[string, string], ctx: ExecutionContex
             firstVal = valueToString(v)
             found = true
             break
-          if found and firstVal == left: return "true"
+          if not found: continue
+          if isNull(firstVal):
+            sawNull = true
+            continue
+          if firstVal == left: return "true"
+        if sawNull: return "\\N"
         return "false"
+      if isNull(right): return "\\N"
       try:
         let lv = parseFloat(left)
         let rv = parseFloat(right)
@@ -524,8 +549,10 @@ proc evalExprOld*(expr: IRExpr, row: Table[string, string], ctx: ExecutionContex
       except CatchableError: discard
       return if left == right: "true" else: "false"
     of irNotIn:
+      if isNull(left): return "\\N"
       if expr.binRight.kind == irekSubquery:
         let subRows = requireExecutePlanHook()(ctx, expr.binRight.subqueryPlan)
+        var sawNull = false
         for row in subRows:
           # Compare against the first non-internal column only (SQL semantics)
           var firstVal = ""
@@ -535,8 +562,14 @@ proc evalExprOld*(expr: IRExpr, row: Table[string, string], ctx: ExecutionContex
             firstVal = valueToString(v)
             found = true
             break
-          if found and firstVal == left: return "false"
+          if not found: continue
+          if isNull(firstVal):
+            sawNull = true
+            continue
+          if firstVal == left: return "false"
+        if sawNull: return "\\N"
         return "true"
+      if isNull(right): return "\\N"
       try:
         let lv = parseFloat(left)
         let rv = parseFloat(right)
@@ -661,6 +694,7 @@ proc evalExprOld*(expr: IRExpr, row: Table[string, string], ctx: ExecutionContex
     case expr.unOp
     of irNot:
       let v = evalExprOld(expr.unExpr, row, ctx)
+      if isNull(v): return "\\N"
       return if v == "true": "false" else: "true"
     of irIsNull:
       let v = evalExprOld(expr.unExpr, row, ctx)
@@ -923,7 +957,6 @@ proc evalExprOld*(expr: IRExpr, row: Table[string, string], ctx: ExecutionContex
         ddl.add("\n")
 
       # Sample data
-      var kvPairs: seq[(string, seq[byte])] = @[]
       let rows = requireExecScanHook()(ctx, table)
       let sampleLimit = min(5, rows.len)
       if sampleLimit > 0:

src/barabadb/query/exec/fk.nim

@@ -30,14 +30,14 @@ proc enforceFkOnDelete*(ctx: ExecutionContext, parentTable: string, parentCol: s
           of "CASCADE":
             for refRow in refs:
               if "$key" in refRow:
-                var dummy: seq[(string, seq[byte])] = @[]
+                var dummy: seq[tuple[key: string, value: seq[byte], deleted: bool]] = @[]
                 discard execDelete(ctx, childTblName, valueToString(refRow["$key"]), dummy)
           of "SET NULL":
             for refRow in refs:
               if "$key" in refRow:
                 var sets = initTable[string, string]()
                 sets[col.name] = "\\N"
-                var dummy: seq[(string, seq[byte])] = @[]
+                var dummy: seq[tuple[key: string, value: seq[byte], deleted: bool]] = @[]
                 discard execUpdateRow(ctx, childTblName, valueToString(refRow["$key"]), sets, dummy)
           of "RESTRICT", "NO ACTION":
             return (false, "FOREIGN KEY violation: row is referenced by " & childTblName & "." & col.name)
@@ -56,14 +56,14 @@ proc enforceFkOnUpdate*(ctx: ExecutionContext, parentTable: string, parentCol: s
               if "$key" in refRow:
                 var sets = initTable[string, string]()
                 sets[col.name] = newVal
-                var dummy: seq[(string, seq[byte])] = @[]
+                var dummy: seq[tuple[key: string, value: seq[byte], deleted: bool]] = @[]
                 discard execUpdateRow(ctx, childTblName, valueToString(refRow["$key"]), sets, dummy)
           of "SET NULL":
             for refRow in refs:
               if "$key" in refRow:
                 var sets = initTable[string, string]()
                 sets[col.name] = "\\N"
-                var dummy: seq[(string, seq[byte])] = @[]
+                var dummy: seq[tuple[key: string, value: seq[byte], deleted: bool]] = @[]
                 discard execUpdateRow(ctx, childTblName, valueToString(refRow["$key"]), sets, dummy)
           of "RESTRICT", "NO ACTION":
             return (false, "FOREIGN KEY violation: row is referenced by " & childTblName & "." & col.name)

src/barabadb/query/exec/lower.nim

@@ -76,6 +76,8 @@ proc lowerExpr*(node: Node): IRExpr =
     of bkJsonContainedBy: irOp = irJsonContainedBy
     of bkJsonHasAny: irOp = irJsonHasAny
     of bkJsonHasAll: irOp = irJsonHasAll
+    of bkPow: irOp = irPow
+    of bkConcat: irOp = irAdd  # irAdd concatenates string operands
     else: irOp = irEq
     result.binOp = irOp
     result.binLeft = lowerExpr(node.binLeft)
@@ -120,6 +122,7 @@ proc lowerExpr*(node: Node): IRExpr =
       else: discard
       result.aggArgs = @[]
       for arg in node.funcArgs: result.aggArgs.add(lowerExpr(arg))
+      result.aggDistinct = node.funcDistinct
       if node.funcFilter != nil:
         result.aggFilter = lowerExpr(node.funcFilter)
     else:
@@ -408,9 +411,18 @@ proc lowerSelect*(node: Node): IRPlan =
   if node.selLimit != nil or node.selOffset != nil:
     let limitPlan = IRPlan(kind: irpkLimit)
     limitPlan.limitSource = result
-    limitPlan.limitCount = if node.selLimit != nil and node.selLimit.limitExpr.kind == nkIntLit:
-      node.selLimit.limitExpr.intVal else: 0
-    limitPlan.limitOffset = if node.selOffset != nil and node.selOffset.offsetExpr.kind == nkIntLit:
-      node.selOffset.offsetExpr.intVal else: 0
+    # limitCount: -1 = unlimited (OFFSET without LIMIT). LIMIT 0 is empty.
+    # Negative LIMIT/OFFSET are clamped so slicing cannot IndexDefect.
+    if node.selLimit != nil:
+      if node.selLimit.limitExpr.kind == nkIntLit:
+        limitPlan.limitCount = max(0'i64, node.selLimit.limitExpr.intVal)
+      else:
+        limitPlan.limitCount = 0
+    else:
+      limitPlan.limitCount = -1
+    if node.selOffset != nil and node.selOffset.offsetExpr.kind == nkIntLit:
+      limitPlan.limitOffset = max(0'i64, node.selOffset.offsetExpr.intVal)
+    else:
+      limitPlan.limitOffset = 0
     result = limitPlan
 

src/barabadb/query/exec/plan_exec.nim

@@ -5,6 +5,7 @@
 ## executor split). Pure code motion — no behavior changes.
 import std/strutils
 import std/tables
+import std/sets
 import std/sequtils
 import std/algorithm
 import ../ir
@@ -19,6 +20,19 @@ import eval
 import scan
 import window
 
+# ----------------------------------------------------------------------
+# Aggregate DISTINCT helpers
+# ----------------------------------------------------------------------
+
+proc shouldKeepDistinct(seen: var HashSet[string], s: string, doDistinct: bool): bool =
+  ## Returns true if `s` should be counted/included (first occurrence when distinct).
+  if not doDistinct:
+    return true
+  if s in seen:
+    return false
+  seen.incl(s)
+  return true
+
 # ----------------------------------------------------------------------
 # IR Plan Execution (with actual filter/sort/projection)
 # ----------------------------------------------------------------------
@@ -113,49 +127,71 @@ proc executePlan*(ctx: ExecutionContext, plan: IRPlan): seq[Row] =
                 newRow[alias] = $filteredRows.len
               else:
                 var count = 0
+                var seen: HashSet[string]
                 for row in filteredRows:
                   let v = evalExpr(expr.aggArgs[0], row, ctx)
-                  if v.kind != vkNull: count += 1
+                  if v.kind != vkNull:
+                    let s = valueToString(v)
+                    if shouldKeepDistinct(seen, s, expr.aggDistinct):
+                      count += 1
                 newRow[alias] = $count
             of irSum:
               var sum = 0.0
+              var seen: HashSet[string]
               for row in filteredRows:
                 let v = evalExpr(expr.aggArgs[0], row, ctx)
-                try: sum += parseFloat(valueToString(v)) except CatchableError: discard
+                let s = valueToString(v)
+                if shouldKeepDistinct(seen, s, expr.aggDistinct):
+                  try: sum += parseFloat(s) except CatchableError: discard
               newRow[alias] = $sum
             of irAvg:
               var sum = 0.0
               var count = 0
+              var seen: HashSet[string]
               for row in filteredRows:
                 let v = evalExpr(expr.aggArgs[0], row, ctx)
-                try: sum += parseFloat(valueToString(v)); count += 1 except CatchableError: discard
+                let s = valueToString(v)
+                if shouldKeepDistinct(seen, s, expr.aggDistinct):
+                  try: sum += parseFloat(s); count += 1 except CatchableError: discard
               newRow[alias] = if count > 0: $(sum / float(count)) else: "0"
             of irMin:
               var minVal = ""
+              var seen: HashSet[string]
               for row in filteredRows:
                 let v = evalExpr(expr.aggArgs[0], row, ctx)
                 if v.kind == vkNull: continue
-                if minVal == "" or cmpMin(valueToString(v), minVal): minVal = valueToString(v)
+                let s = valueToString(v)
+                if shouldKeepDistinct(seen, s, expr.aggDistinct):
+                  if minVal == "" or cmpMin(s, minVal): minVal = s
               newRow[alias] = minVal
             of irMax:
               var maxVal = ""
+              var seen: HashSet[string]
               for row in filteredRows:
                 let v = evalExpr(expr.aggArgs[0], row, ctx)
                 if v.kind == vkNull: continue
-                if maxVal == "" or cmpMax(valueToString(v), maxVal): maxVal = valueToString(v)
+                let s = valueToString(v)
+                if shouldKeepDistinct(seen, s, expr.aggDistinct):
+                  if maxVal == "" or cmpMax(s, maxVal): maxVal = s
               newRow[alias] = maxVal
             of irArrayAgg:
               var arr: seq[string]
+              var seen: HashSet[string]
               for row in filteredRows:
                 if expr.aggArgs.len > 0:
-                  arr.add(valueToString(evalExpr(expr.aggArgs[0], row, ctx)))
+                  let s = valueToString(evalExpr(expr.aggArgs[0], row, ctx))
+                  if shouldKeepDistinct(seen, s, expr.aggDistinct):
+                    arr.add(s)
               newRow[alias] = "[" & arr.join(", ") & "]"
             of irStringAgg:
               var parts: seq[string]
+              var seen: HashSet[string]
               let delim = if expr.aggArgs.len > 1: evalExpr(expr.aggArgs[1], initTable[string, Value](), ctx) else: Value(kind: vkString, strVal: ",")
               for row in filteredRows:
                 if expr.aggArgs.len > 0:
-                  parts.add(valueToString(evalExpr(expr.aggArgs[0], row, ctx)))
+                  let s = valueToString(evalExpr(expr.aggArgs[0], row, ctx))
+                  if shouldKeepDistinct(seen, s, expr.aggDistinct):
+                    parts.add(s)
               newRow[alias] = parts.join(valueToString(delim))
           else:
             let val = evalExpr(expr, if sourceRows.len > 0: sourceRows[0] else: initTable[string, Value](), ctx)
@@ -219,12 +255,18 @@ proc executePlan*(ctx: ExecutionContext, plan: IRPlan): seq[Row] =
   of irpkLimit:
     let sourceRows = executePlan(ctx, plan.limitSource)
     var start = int(plan.limitOffset)
+    if start < 0: start = 0
     if start > sourceRows.len: start = sourceRows.len
+    if plan.limitCount < 0:
+      # OFFSET without LIMIT — return the remainder.
+      return sourceRows[start ..< sourceRows.len]
     if plan.limitCount == 0:
       return @[]
     var endIdx = start + int(plan.limitCount)
     if endIdx > sourceRows.len:
       endIdx = sourceRows.len
+    if endIdx < start:
+      endIdx = start
     return sourceRows[start..<endIdx]
 
   of irpkGroupBy:
@@ -292,49 +334,71 @@ proc executePlan*(ctx: ExecutionContext, plan: IRPlan): seq[Row] =
               aggRow[aggKey] = $filteredRows.len
             else:
               var count = 0
+              var seen: HashSet[string]
               for row in filteredRows:
                 let v = evalExpr(aggExpr.aggArgs[0], row, ctx)
-                if v.kind != vkNull: count += 1
+                if v.kind != vkNull:
+                  let s = valueToString(v)
+                  if shouldKeepDistinct(seen, s, aggExpr.aggDistinct):
+                    count += 1
               aggRow[aggKey] = $count
           of irSum:
             var sum = 0.0
+            var seen: HashSet[string]
             for row in filteredRows:
               let v = evalExpr(aggExpr.aggArgs[0], row, ctx)
-              try: sum += parseFloat(valueToString(v)) except CatchableError: discard
+              let s = valueToString(v)
+              if shouldKeepDistinct(seen, s, aggExpr.aggDistinct):
+                try: sum += parseFloat(s) except CatchableError: discard
             aggRow[aggKey] = $sum
           of irAvg:
             var sum = 0.0
             var count = 0
+            var seen: HashSet[string]
             for row in filteredRows:
               let v = evalExpr(aggExpr.aggArgs[0], row, ctx)
-              try: sum += parseFloat(valueToString(v)); count += 1 except CatchableError: discard
+              let s = valueToString(v)
+              if shouldKeepDistinct(seen, s, aggExpr.aggDistinct):
+                try: sum += parseFloat(s); count += 1 except CatchableError: discard
             aggRow[aggKey] = if count > 0: $(sum / float(count)) else: "0"
           of irMin:
             var minVal = ""
+            var seen: HashSet[string]
             for row in filteredRows:
               let v = evalExpr(aggExpr.aggArgs[0], row, ctx)
               if v.kind == vkNull: continue
-              if minVal == "" or cmpMin(valueToString(v), minVal): minVal = valueToString(v)
+              let s = valueToString(v)
+              if shouldKeepDistinct(seen, s, aggExpr.aggDistinct):
+                if minVal == "" or cmpMin(s, minVal): minVal = s
             aggRow[aggKey] = minVal
           of irMax:
             var maxVal = ""
+            var seen: HashSet[string]
             for row in filteredRows:
               let v = evalExpr(aggExpr.aggArgs[0], row, ctx)
               if v.kind == vkNull: continue
-              if maxVal == "" or cmpMax(valueToString(v), maxVal): maxVal = valueToString(v)
+              let s = valueToString(v)
+              if shouldKeepDistinct(seen, s, aggExpr.aggDistinct):
+                if maxVal == "" or cmpMax(s, maxVal): maxVal = s
             aggRow[aggKey] = maxVal
           of irArrayAgg:
             var arr: seq[string]
+            var seen: HashSet[string]
             for row in filteredRows:
               if aggExpr.aggArgs.len > 0:
-                arr.add(valueToString(evalExpr(aggExpr.aggArgs[0], row, ctx)))
+                let s = valueToString(evalExpr(aggExpr.aggArgs[0], row, ctx))
+                if shouldKeepDistinct(seen, s, aggExpr.aggDistinct):
+                  arr.add(s)
             aggRow[aggKey] = "[" & arr.join(", ") & "]"
           of irStringAgg:
             var parts: seq[string]
+            var seen: HashSet[string]
             let delim = if aggExpr.aggArgs.len > 1: evalExpr(aggExpr.aggArgs[1], initTable[string, Value](), ctx) else: Value(kind: vkString, strVal: ",")
             for row in filteredRows:
               if aggExpr.aggArgs.len > 0:
-                parts.add(valueToString(evalExpr(aggExpr.aggArgs[0], row, ctx)))
+                let s = valueToString(evalExpr(aggExpr.aggArgs[0], row, ctx))
+                if shouldKeepDistinct(seen, s, aggExpr.aggDistinct):
+                  parts.add(s)
             aggRow[aggKey] = parts.join(valueToString(delim))
         # Apply HAVING filter
         if plan.groupHaving != nil:

src/barabadb/query/exec/rls.nim

@@ -11,20 +11,22 @@ import lower
 # Row-Level Security
 # ----------------------------------------------------------------------
 
-proc hasPrivilege*(ctx: ExecutionContext, tableName, command: string): bool =
-  if ctx.currentUser.len == 0: return true
-  let user = ctx.users.getOrDefault(ctx.currentUser)
+proc hasPrivilegeFor*(ctx: ExecutionContext, username, tableName, command: string): bool =
+  ## Privilege check for an explicit username (does not mutate ctx.currentUser).
+  if username.len == 0: return true
+  let user = ctx.users.getOrDefault(username)
   if user.isSuperuser: return true
-  # Check table-level policies for user or PUBLIC
-  # For now: if no policies exist, allow everything (backward compatible)
   if tableName notin ctx.policies: return true
   let policies = ctx.policies[tableName]
-  # If RLS is enabled (policies exist), check if user matches any policy
   for pol in policies:
     if pol.command == "ALL" or pol.command == command:
       return true
   return false
 
+proc hasPrivilege*(ctx: ExecutionContext, tableName, command: string): bool =
+  if ctx.currentUser.len == 0: return true
+  hasPrivilegeFor(ctx, ctx.currentUser, tableName, command)
+
 proc passesPolicy*(ctx: ExecutionContext, tableName, command: string, row: Row): bool =
   if ctx.currentUser.len == 0: return true
   let user = ctx.users.getOrDefault(ctx.currentUser)

src/barabadb/query/exec/types.nim

@@ -136,13 +136,13 @@ type
     rows*: seq[Row]
     affectedRows*: int
     message*: string
-    keyValuePairs*: seq[(string, seq[byte])]
+    keyValuePairs*: seq[tuple[key: string, value: seq[byte], deleted: bool]]
 
 proc `==`*(a, b: IndexEntry): bool =
   a.lsmKey == b.lsmKey and a.rowValue == b.rowValue
 
 proc okResult*(rows: seq[Row] = @[], cols: seq[string] = @[], affected: int = 0, msg: string = "",
-               kvPairs: seq[(string, seq[byte])] = @[]): ExecResult =
+               kvPairs: seq[tuple[key: string, value: seq[byte], deleted: bool]] = @[]): ExecResult =
   ExecResult(success: true, columns: cols, rows: rows, affectedRows: affected, message: msg,
              keyValuePairs: kvPairs)
 

src/barabadb/query/exec/window.nim

@@ -177,6 +177,64 @@ proc computeWindowValues*(rows: seq[Row], expr: IRExpr, ctx: ExecutionContext =
       for pos, rowIdx in sortedIdxs:
         let (_, fEnd) = resolveFrameBounds(pos, sortedIdxs.len, frameStart, frameEnd)
         result[rowIdx] = valueToString(evalExpr(expr.wfArgs[0], rows[sortedIdxs[fEnd]], ctx))
+    of "sum", "avg", "count", "min", "max":
+      let countAll = wfName == "count" and
+        (expr.wfArgs.len == 0 or expr.wfArgs[0].kind == irekStar)
+      for pos, rowIdx in sortedIdxs:
+        let (fStart, fEnd) = resolveFrameBounds(pos, sortedIdxs.len, frameStart, frameEnd)
+        if wfName == "count" and countAll:
+          result[rowIdx] = $(fEnd - fStart + 1)
+          continue
+        if wfName == "count":
+          var cnt = 0
+          if expr.wfArgs.len > 0:
+            for i in fStart .. fEnd:
+              let s = valueToString(evalExpr(expr.wfArgs[0], rows[sortedIdxs[i]], ctx))
+              if not isNull(s) and s.len > 0:
+                inc cnt
+          result[rowIdx] = $cnt
+          continue
+        if expr.wfArgs.len == 0:
+          result[rowIdx] = "\\N"
+          continue
+        var sum = 0.0
+        var cnt = 0
+        var minF = 0.0
+        var maxF = 0.0
+        var minS = ""
+        var maxS = ""
+        var allNumeric = true
+        for i in fStart .. fEnd:
+          let s = valueToString(evalExpr(expr.wfArgs[0], rows[sortedIdxs[i]], ctx))
+          if isNull(s) or s.len == 0: continue
+          inc cnt
+          if cnt == 1 or s < minS: minS = s
+          if cnt == 1 or s > maxS: maxS = s
+          try:
+            let f = parseFloat(s)
+            sum += f
+            if cnt == 1:
+              minF = f
+              maxF = f
+            else:
+              if f < minF: minF = f
+              if f > maxF: maxF = f
+          except CatchableError:
+            allNumeric = false
+        if cnt == 0:
+          result[rowIdx] = "\\N"
+        else:
+          case wfName
+          of "sum":
+            result[rowIdx] = if allNumeric: $sum else: "\\N"
+          of "avg":
+            result[rowIdx] = if allNumeric: $(sum / float(cnt)) else: "\\N"
+          of "min":
+            result[rowIdx] = if allNumeric: $minF else: minS
+          of "max":
+            result[rowIdx] = if allNumeric: $maxF else: maxS
+          else:
+            result[rowIdx] = "\\N"
     else:
       # Unknown window function — fill with null
       for rowIdx in sortedIdxs:

src/barabadb/query/executor.nim

@@ -448,6 +448,26 @@ proc executeQueryImpl(ctx: ExecutionContext, astNode: Node, params: seq[WireValu
     if cols.len == 0:
       cols = rightRes.columns
 
+    # Fingerprint a projected row for set-op dedup. Prefer declared columns;
+    # fall back to non-system keys so UNION/INTERSECT/EXCEPT work without `$value`.
+    proc setOpRowKey(row: Row, colNames: seq[string]): string =
+      var parts: seq[string] = @[]
+      if colNames.len > 0:
+        for c in colNames:
+          if c in row:
+            parts.add(valueToString(row[c]))
+          else:
+            parts.add("")
+      else:
+        var keys: seq[string] = @[]
+        for k, _ in row:
+          if not k.startsWith("$"):
+            keys.add(k)
+        keys.sort()
+        for k in keys:
+          parts.add(k & "=" & valueToString(row[k]))
+      return parts.join("\x1f")
+
     var rows: seq[Row] = @[]
     case stmt.setOpKind
     of sdkUnion:
@@ -460,28 +480,30 @@ proc executeQueryImpl(ctx: ExecutionContext, astNode: Node, params: seq[WireValu
         # UNION: deduplicate
         var seen: Table[string, bool]
         for row in leftRes.rows:
-          seen[valueToString(row["$value"])] = true
+          seen[setOpRowKey(row, cols)] = true
         for row in rightRes.rows:
-          if not seen.getOrDefault(valueToString(row["$value"]), false):
-            seen[valueToString(row["$value"])] = true
+          let k = setOpRowKey(row, cols)
+          if not seen.getOrDefault(k, false):
+            seen[k] = true
             rows.add(row)
 
     of sdkIntersect:
       var leftSet: Table[string, bool]
       for row in leftRes.rows:
-        leftSet[valueToString(row["$value"])] = true
+        leftSet[setOpRowKey(row, cols)] = true
       for row in rightRes.rows:
-        if leftSet.getOrDefault(valueToString(row["$value"]), false):
+        let k = setOpRowKey(row, cols)
+        if leftSet.getOrDefault(k, false):
           rows.add(row)
           if not stmt.setOpAll:
-            leftSet.del(valueToString(row["$value"]))  # remove to prevent duplicates for INTERSECT (not ALL)
+            leftSet.del(k)  # remove to prevent duplicates for INTERSECT (not ALL)
 
     of sdkExcept:
       var rightSet: Table[string, bool]
       for row in rightRes.rows:
-        rightSet[valueToString(row["$value"])] = true
+        rightSet[setOpRowKey(row, cols)] = true
       for row in leftRes.rows:
-        if not rightSet.getOrDefault(valueToString(row["$value"]), false):
+        if not rightSet.getOrDefault(setOpRowKey(row, cols), false):
           rows.add(row)
 
     return okResult(rows, cols)
@@ -574,7 +596,7 @@ proc executeQueryImpl(ctx: ExecutionContext, astNode: Node, params: seq[WireValu
           row[f] = mutableValues[0][i]
     fireTriggers(ctx, stmt.insTarget, "before", "insert", row)
 
-    var kvPairs: seq[(string, seq[byte])]
+    var kvPairs: seq[tuple[key: string, value: seq[byte], deleted: bool]]
     let count = execInsert(ctx, stmt.insTarget, mutableFields, mutableValues, kvPairs)
 
     # Fire AFTER INSERT triggers
@@ -626,7 +648,7 @@ proc executeQueryImpl(ctx: ExecutionContext, astNode: Node, params: seq[WireValu
     # Scan and apply
     let rows = execScan(ctx, stmt.updTarget)
     var count = 0
-    var kvPairs: seq[(string, seq[byte])]
+    var kvPairs: seq[tuple[key: string, value: seq[byte], deleted: bool]]
     for row in rows:
       # Compute sets for this row (expressions may reference columns)
       var sets = initTable[string, string]()
@@ -701,7 +723,7 @@ proc executeQueryImpl(ctx: ExecutionContext, astNode: Node, params: seq[WireValu
     # Delete all rows matching WHERE
     let rows = execScan(ctx, stmt.delTarget)
     var count = 0
-    var kvPairs: seq[(string, seq[byte])]
+    var kvPairs: seq[tuple[key: string, value: seq[byte], deleted: bool]]
     for row in rows:
       if stmt.delWhere != nil and stmt.delWhere.whereExpr != nil:
         let whereExpr = lowerExpr(stmt.delWhere.whereExpr)
@@ -744,7 +766,7 @@ proc executeQueryImpl(ctx: ExecutionContext, astNode: Node, params: seq[WireValu
 
     let targetRows = execScan(ctx, stmt.mergeTarget)
     var count = 0
-    var kvPairs: seq[(string, seq[byte])]
+    var kvPairs: seq[tuple[key: string, value: seq[byte], deleted: bool]]
 
     for srcRow in sourceRows:
       var matched = false
@@ -759,21 +781,34 @@ proc executeQueryImpl(ctx: ExecutionContext, astNode: Node, params: seq[WireValu
         let onExpr = lowerExpr(stmt.mergeOn)
         if valueToString(evalExpr(onExpr, rowWithTarget, ctx)) == "true":
           matched = true
-          if stmt.mergeMatchedUpdate.len > 0 and "$key" in tgtRow:
-            var updateSets = initTable[string, string]()
-            for s in stmt.mergeMatchedUpdate:
-              if s.kind == nkBinOp and s.binOp == bkAssign:
-                if s.binLeft.kind == nkIdent:
-                  let valExpr = lowerExpr(s.binRight)
-                  updateSets[s.binLeft.identName] = valueToString(evalExpr(valExpr, rowWithTarget, ctx))
-            var newRow = tgtRow
-            for col, val in updateSets:
-              newRow[col] = Value(kind: vkString, strVal: val)
-            fireTriggers(ctx, stmt.mergeTarget, "before", "update", tgtRow)
-            count += execUpdateRow(ctx, stmt.mergeTarget, valueToString(tgtRow["$key"]), updateSets, kvPairs)
-            fireTriggers(ctx, stmt.mergeTarget, "after", "update", newRow)
-            if ctx.onChange != nil:
-              ctx.onChange(ChangeEvent(table: stmt.mergeTarget, kind: ckUpdate, key: valueToString(tgtRow["$key"]), data: ""))
+          # Optional AND <condition> after WHEN MATCHED
+          var applyMatched = true
+          if stmt.mergeMatchedCondition != nil:
+            let condExpr = lowerExpr(stmt.mergeMatchedCondition)
+            applyMatched = valueToString(evalExpr(condExpr, rowWithTarget, ctx)) == "true"
+          if applyMatched and "$key" in tgtRow:
+            if stmt.mergeMatchedDelete:
+              fireTriggers(ctx, stmt.mergeTarget, "before", "delete", tgtRow)
+              count += execDelete(ctx, stmt.mergeTarget, valueToString(tgtRow["$key"]), kvPairs)
+              fireTriggers(ctx, stmt.mergeTarget, "after", "delete", tgtRow)
+              if ctx.onChange != nil:
+                ctx.onChange(ChangeEvent(table: stmt.mergeTarget, kind: ckDelete,
+                  key: valueToString(tgtRow["$key"]), data: ""))
+            elif stmt.mergeMatchedUpdate.len > 0:
+              var updateSets = initTable[string, string]()
+              for s in stmt.mergeMatchedUpdate:
+                if s.kind == nkBinOp and s.binOp == bkAssign:
+                  if s.binLeft.kind == nkIdent:
+                    let valExpr = lowerExpr(s.binRight)
+                    updateSets[s.binLeft.identName] = valueToString(evalExpr(valExpr, rowWithTarget, ctx))
+              var newRow = tgtRow
+              for col, val in updateSets:
+                newRow[col] = Value(kind: vkString, strVal: val)
+              fireTriggers(ctx, stmt.mergeTarget, "before", "update", tgtRow)
+              count += execUpdateRow(ctx, stmt.mergeTarget, valueToString(tgtRow["$key"]), updateSets, kvPairs)
+              fireTriggers(ctx, stmt.mergeTarget, "after", "update", newRow)
+              if ctx.onChange != nil:
+                ctx.onChange(ChangeEvent(table: stmt.mergeTarget, kind: ckUpdate, key: valueToString(tgtRow["$key"]), data: ""))
           break
 
       if not matched and stmt.mergeNotMatchedInsert.len > 0:
@@ -793,7 +828,7 @@ proc executeQueryImpl(ctx: ExecutionContext, astNode: Node, params: seq[WireValu
           for i, f in fields:
             if i < values.len: row[f] = Value(kind: vkString, strVal: values[i])
           fireTriggers(ctx, stmt.mergeTarget, "before", "insert", row)
-          var insKvPairs: seq[(string, seq[byte])]
+          var insKvPairs: seq[tuple[key: string, value: seq[byte], deleted: bool]]
           count += execInsert(ctx, stmt.mergeTarget, fields, @[values], insKvPairs)
           for kv in insKvPairs: kvPairs.add(kv)
           fireTriggers(ctx, stmt.mergeTarget, "after", "insert", row)
@@ -982,16 +1017,17 @@ proc executeQueryImpl(ctx: ExecutionContext, astNode: Node, params: seq[WireValu
 
   of nkCommitTxn:
     if ctx.pendingTxn != nil and ctx.pendingTxn.state == tsActive:
-      var kvPairs: seq[(string, seq[byte])]
+      var kvPairs: seq[tuple[key: string, value: seq[byte], deleted: bool]]
       for key, version in ctx.pendingTxn.writeSet:
         if version.isDelete:
           ctx.db.delete(key)
-          # Empty value is the raft/replication "delete" convention — never
-          # ship a non-empty body for isDelete or followers will resurrect.
-          kvPairs.add((key, @[]))
+          # Empty value + deleted=true is the raft/replication "delete"
+          # convention — never ship a non-empty body for isDelete or
+          # followers will resurrect.
+          kvPairs.add((key, @[], true))
         else:
           ctx.db.put(key, version.value)
-          kvPairs.add((key, version.value))
+          kvPairs.add((key, version.value, false))
       discard ctx.txnManager.commit(ctx.pendingTxn)
       ctx.pendingTxn = nil
       return okResult(msg="Transaction committed", kvPairs=kvPairs)

src/barabadb/query/parser.nim

@@ -194,6 +194,11 @@ proc parsePrimary(p: var Parser): Node =
       discard p.expect(tkWhere)
       node.funcFilter = p.parseExpr()
       discard p.expect(tkRParen)
+    # Window aggregate: SUM/AVG/COUNT/MIN/MAX(...) OVER (...)
+    if p.peek().kind == tkOver:
+      let overClause = p.parseOverClause()
+      return Node(kind: nkWindowExpr, winFunc: funcName.toLower(), winArgs: args,
+                  winOver: overClause, line: tok.line, col: tok.col)
     return node
   of tkCase:
     discard p.advance()

src/barabadb/storage/btree.nim

@@ -175,6 +175,61 @@ proc scan*[K, V](btree: BTreeIndex[K, V], startKey, endKey: K): seq[(K, seq[V])]
   finally:
     release(btree.lock)
 
+proc subtreeMinMax[K, V](node: BTreeNode[K, V]): (bool, K, K) =
+  ## Inclusive min/max of keys stored in this subtree's leaves.
+  if node == nil:
+    return (false, default(K), default(K))
+  if node.isLeaf:
+    if node.keys.len == 0:
+      return (false, default(K), default(K))
+    return (true, node.keys[0], node.keys[^1])
+  var have = false
+  var mn, mx: K
+  for c in node.children:
+    let (ok, a, b) = subtreeMinMax(c)
+    if ok:
+      if not have:
+        mn = a
+        mx = b
+        have = true
+      else:
+        if a < mn: mn = a
+        if b > mx: mx = b
+  return (have, mn, mx)
+
+proc collectSeparatorErrors[K, V](node: BTreeNode[K, V], errors: var seq[string]) =
+  ## Search uses `key > separator → right child`, so every key in the left
+  ## subtree must be <= sep (otherwise it is routed right and missed — no
+  ## prev-leaf pointer). Boundary duplicates are allowed: the same key may be
+  ## max(left) and min(right); get/remove walk `next` to collect them.
+  if node == nil or node.isLeaf:
+    return
+  for i in 0..<node.keys.len:
+    if i + 1 >= node.children.len:
+      errors.add("internal node has fewer children than keys+1")
+      break
+    let (lok, _, lmax) = subtreeMinMax(node.children[i])
+    let (rok, rmin, _) = subtreeMinMax(node.children[i + 1])
+    let sep = node.keys[i]
+    if lok and lmax > sep:
+      errors.add("separator[" & $i & "]=" & $sep &
+        " < max(left)=" & $lmax & " (search would miss left keys)")
+    if lok and rok and rmin < lmax:
+      errors.add("separator[" & $i & "]=" & $sep &
+        " leaf order inverted: max(left)=" & $lmax & " > min(right)=" & $rmin)
+    collectSeparatorErrors(node.children[i], errors)
+  if node.children.len > 0:
+    collectSeparatorErrors(node.children[^1], errors)
+
+proc checkInvariants*[K, V](btree: BTreeIndex[K, V]): seq[string] =
+  ## Returns a list of separator/search-routing violations (empty = healthy).
+  acquire(btree.lock)
+  try:
+    result = @[]
+    collectSeparatorErrors(btree.root, result)
+  finally:
+    release(btree.lock)
+
 proc len*[K, V](btree: BTreeIndex[K, V]): int =
   acquire(btree.lock)
   try:
@@ -209,7 +264,9 @@ proc borrowFromLeft[K, V](node: BTreeNode[K, V], parent: BTreeNode[K, V], parent
     node.values.insert(borrowVal, 0)
     sibling.keys.setLen(sibling.keys.len - 1)
     sibling.values.setLen(sibling.values.len - 1)
-    parent.keys[parentIdx - 1] = node.keys[0]
+    # Search is `key > sep → right` (left-max). After lending, sep is the
+    # left sibling's new max — not the borrowed key now sitting in `node`.
+    parent.keys[parentIdx - 1] = sibling.keys[^1]
   else:
     # Borrow from internal sibling
     let borrowKey = sibling.keys[^1]
@@ -231,7 +288,8 @@ proc borrowFromRight[K, V](node: BTreeNode[K, V], parent: BTreeNode[K, V], paren
     node.values.add(borrowVal)
     sibling.keys.delete(0)
     sibling.values.delete(0)
-    parent.keys[parentIdx] = sibling.keys[0]
+    # Borrowed key is now left's max; search must keep it on the left.
+    parent.keys[parentIdx] = node.keys[^1]
   else:
     let borrowKey = sibling.keys[0]
     let borrowChild = sibling.children[0]
@@ -325,10 +383,14 @@ proc rebalanceAfterDelete[K, V](node: BTreeNode[K, V], root: var BTreeNode[K, V]
     mergeWithLeft(node, parent, parentIdx)
   elif hasRight:
     mergeWithRight(node, parent, parentIdx)
+  else:
+    return
 
-  # Recursively rebalance parent if it fell below minimum
+  # Merge dropped a separator from parent — rebalance up the tree.
   if parent == root and parent.keys.len == 0 and parent.children.len == 1:
     root = parent.children[0]
+  elif parent.keys.len < minKeysForLeaf(parent, order):
+    rebalanceAfterDelete(parent, root, order)
 
 proc remove*[K, V](btree: var BTreeIndex[K, V], key: K, value: V) =
   acquire(btree.lock)
@@ -368,15 +430,17 @@ proc remove*[K, V](btree: var BTreeIndex[K, V], key: K, value: V) =
       else:
         # Internal node: recurse into child
         let child = node.children[i]
-        let oldFirstKey = if child.keys.len > 0: child.keys[0] else: default(K)
         let found = removeRec(child, root, order)
         if found:
-          # Update separator if child's first key changed.
-          # Separator node.keys[i-1] represents child's first key (for i > 0).
-          if i > 0 and child.keys.len > 0 and child.keys[0] != oldFirstKey:
-            node.keys[i - 1] = child.keys[0]
-          # Rebalance the child if needed
+          # Rebalance first — merge/borrow rewrite parent separators.
           rebalanceAfterDelete(child, root, order)
+          # Leaf children only: refresh left-max separator. Skip if merge
+          # already unlinked this child. Internal keys are promoted
+          # separators, not copies of child.keys[^1]; copying those (the
+          # naive H10 rewrite) makes sep < max(left) and misses keys.
+          if i < node.children.len and node.children[i] == child and
+             child.isLeaf and child.keys.len > 0 and i < node.keys.len:
+            node.keys[i] = child.keys[^1]
         return found
 
     if removeRec(btree.root, btree.root, btree.order):

src/barabadb/storage/compaction.nim

@@ -125,13 +125,16 @@ proc compact*(cs: CompactionStrategy, level: int): CompactionResult =
     return cmp(b.timestamp, a.timestamp)  # newest first
   )
 
-  # Deduplicate: keep only the newest version of each key
+  # Deduplicate: keep only the newest version of each key.
+  # Use a haveLast flag — sentinel lastKey="" would skip the empty-string key.
   var merged: seq[Entry] = @[]
   var lastKey = ""
+  var haveLast = false
   for entry in allEntries:
-    if entry.key != lastKey:
+    if not haveLast or entry.key != lastKey:
       merged.add(entry)
       lastKey = entry.key
+      haveLast = true
 
   # Keep tombstones to prevent deleted keys from resurrecting in lower levels
   var final: seq[Entry] = @[]
@@ -155,20 +158,15 @@ proc compact*(cs: CompactionStrategy, level: int): CompactionResult =
     createdAt: tables[^1].createdAt,
   )
 
-  # Verify output SSTable before deleting sources
+  # Verify output SSTable before mutating the catalog. Input files stay on
+  # disk until the caller loads the output and writes MANIFEST (crash-safe
+  # order: output durable + catalog updated, then unlink inputs).
   let (ok, msg) = verifySSTable(outputPath)
   if not ok:
     echo "[ERROR] Compaction output verification failed: ", msg
     try: removeFile(outputPath) except CatchableError: discard
     return CompactionResult()
 
-  # Remove old SSTable files
-  for t in tables:
-    try:
-      removeFile(t.path)
-    except CatchableError as e:
-      echo "[WARN] Failed to remove old SSTable: ", t.path, ": ", e.msg
-
   # Update level arrays
   var newTables: seq[SSTableMeta] = @[]
   for t in cs.levels[level]:

src/barabadb/storage/lsm.nim

@@ -696,6 +696,10 @@ proc newLSMTree*(
         var version: uint32 = 0
         if stream.readData(addr magic, 4) == 4 and magic == WALMagic:
           if stream.readData(addr version, 4) == 4:
+            # Cap per-record sizes to avoid multi-GiB alloc on torn/corrupt WAL.
+            # Kind must be a known WalEntryKind value (1..4) — out-of-range casts
+            # raise CaseStmtError (Defect) and crash the process.
+            const MaxWalRecordField = 64 * 1024 * 1024  # 64 MB
             while not stream.atEnd():
               var kind: uint8 = 0
               var timestamp: uint64 = 0
@@ -704,10 +708,20 @@ proc newLSMTree*(
               if stream.readData(addr kind, 1) != 1: break
               if stream.readData(addr timestamp, 8) != 8: break
               if stream.readData(addr keyLen, 4) != 4: break
+              if keyLen.int > MaxWalRecordField:
+                echo "[WARN] WAL recovery: torn/corrupt record (keyLen=", keyLen, ") — stopping replay"
+                break
+              # Validate kind before allocating or branching (avoids CaseStmtError Defect)
+              if kind < uint8(wekPut) or kind > uint8(wekCommit):
+                echo "[WARN] WAL recovery: invalid entry kind ", kind, " — stopping replay"
+                break
               var key = newString(keyLen.int)
               if keyLen > 0:
                 if stream.readData(addr key[0], keyLen.int) != keyLen.int: break
               if stream.readData(addr valLen, 4) != 4: break
+              if valLen.int > MaxWalRecordField:
+                echo "[WARN] WAL recovery: torn/corrupt record (valLen=", valLen, ") — stopping replay"
+                break
               var value = newSeq[byte](valLen.int)
               if valLen > 0:
                 if stream.readData(addr value[0], valLen.int) != valLen.int: break
@@ -866,26 +880,36 @@ proc flushUnsafe(db: LSMTree) =
   if db.immutableMem.len == 0 and db.memTable.len == 0:
     return
 
-  # Flush immutable memtable if present, otherwise flush current memtable
-  var toFlush = db.immutableMem
-  if toFlush.len == 0:
-    toFlush = db.memTable
-    db.memTable = newMemTable(db.memMaxSize)
+  # Flush immutable memtable if present, otherwise flush current memtable.
+  # Do NOT clear the source memtable until the SSTable is written — an IOError
+  # mid-write must leave the data still visible to live reads (WAL still has it).
+  var flushingImmutable = false
+  var toFlush: MemTable
+  if db.immutableMem.len > 0:
+    toFlush = db.immutableMem
+    flushingImmutable = true
   else:
-    db.immutableMem = newMemTable(0)
+    toFlush = db.memTable
 
   if toFlush.len == 0:
     return
 
   let path = db.dir / "sstables" / ($db.nextSSTableId & ".sst")
+  let sstId = db.nextSSTableId
   inc db.nextSSTableId
 
   # Sort once at flush time (O(n log n)) — put/get stay O(1)
   var sst = writeSSTable(toFlush.sortedEntries(), path, level = 0)
-  sst.id = db.nextSSTableId - 1
+  sst.id = sstId
   db.sstables.add(sst)
   # SSTables are kept in insertion order (newest last) so getUnsafe can search newest-first
 
+  # Only now drop the in-memory copy — SSTable is durable on disk
+  if flushingImmutable:
+    db.immutableMem = newMemTable(0)
+  else:
+    db.memTable = newMemTable(db.memMaxSize)
+
   # Update MANIFEST atomically
   inc db.manifestSequence
   try:
@@ -930,27 +954,29 @@ proc checkpoint*(db: LSMTree) =
   ## rotate WAL, and write MANIFEST. This provides a clean boundary
   ## for online backup without stopping the server.
   acquireWrite(db.lock)
+  try:
+    # Flush any pending immutable memtable first
+    if db.immutableMem.len > 0:
+      flushUnsafe(db)
 
-  # Flush any pending immutable memtable first
-  if db.immutableMem.len > 0:
-    flushUnsafe(db)
+    # Freeze current memtable so writes can continue on a new one
+    if db.memTable.len > 0:
+      db.immutableMem = db.memTable
+      db.memTable = newMemTable(db.memMaxSize)
 
-  # Freeze current memtable so writes can continue on a new one
-  if db.memTable.len > 0:
-    db.immutableMem = db.memTable
-    db.memTable = newMemTable(db.memMaxSize)
+    # Flush the frozen memtable
+    if db.immutableMem.len > 0:
+      flushUnsafe(db)
 
-  # Flush the frozen memtable
-  if db.immutableMem.len > 0:
-    flushUnsafe(db)
-
-  # Rotate WAL for a clean backup boundary
-  acquire(db.walLock)
-  db.wal.maybeRotate()
-  db.wal.sync()
-  release(db.walLock)
-
-  releaseWrite(db.lock)
+    # Rotate WAL for a clean backup boundary
+    acquire(db.walLock)
+    try:
+      db.wal.maybeRotate()
+      db.wal.sync()
+    finally:
+      release(db.walLock)
+  finally:
+    releaseWrite(db.lock)
 
 proc close*(db: LSMTree) =
   acquireWrite(db.lock)

src/barabadb/storage/mmap.nim

@@ -80,7 +80,8 @@ proc readAt*(mf: MmapFile, offset: int, size: int): seq[byte] =
   if mf.regions.len == 0:
     return @[]
   let region = mf.regions[0]
-  if offset < 0 or size < 0 or offset + size > region.size:
+  # overflow-safe bound: offset > size - length (not offset + length > size)
+  if offset < 0 or size < 0 or size > region.size or offset > region.size - size:
     return @[]
   result = newSeq[byte](size)
   copyMem(addr result[0], unsafeAddr region.data[offset], size)
@@ -91,21 +92,24 @@ proc readByte*(mf: MmapFile, offset: int): byte =
   return mf.regions[0].data[offset]
 
 proc readUint32*(mf: MmapFile, offset: int): uint32 =
-  if mf.regions.len == 0 or offset < 0 or offset + 4 > mf.regions[0].size:
+  if mf.regions.len == 0 or offset < 0 or 4 > mf.regions[0].size or
+     offset > mf.regions[0].size - 4:
     return 0
   var val: uint32
   copyMem(addr val, unsafeAddr mf.regions[0].data[offset], 4)
   return val
 
 proc readUint64*(mf: MmapFile, offset: int): uint64 =
-  if mf.regions.len == 0 or offset < 0 or offset + 8 > mf.regions[0].size:
+  if mf.regions.len == 0 or offset < 0 or 8 > mf.regions[0].size or
+     offset > mf.regions[0].size - 8:
     return 0
   var val: uint64
   copyMem(addr val, unsafeAddr mf.regions[0].data[offset], 8)
   return val
 
 proc readString*(mf: MmapFile, offset: int, size: int): string =
-  if mf.regions.len == 0 or offset < 0 or size < 0 or offset + size > mf.regions[0].size:
+  if mf.regions.len == 0 or offset < 0 or size < 0 or
+     size > mf.regions[0].size or offset > mf.regions[0].size - size:
     return ""
   result = newString(size)
   copyMem(addr result[0], unsafeAddr mf.regions[0].data[offset], size)

src/barabadb/storage/recovery.nim

@@ -68,6 +68,7 @@ proc scanWAL*(rec: CrashRecovery): seq[RecoveredEntry] =
   var txnId: uint64 = 0
   var entryCount = 0
 
+  const MaxWalRecordField = 64 * 1024 * 1024  # 64 MB
   while not stream.atEnd():
     var kind: uint8 = 0
     var timestamp: uint64 = 0
@@ -77,12 +78,15 @@ proc scanWAL*(rec: CrashRecovery): seq[RecoveredEntry] =
     if stream.readData(addr kind, 1) != 1: break
     if stream.readData(addr timestamp, 8) != 8: break
     if stream.readData(addr keyLen, 4) != 4: break
+    if keyLen.int > MaxWalRecordField: break
+    if kind < uint8(wekPut) or kind > uint8(wekCommit): break
 
     var key = newString(keyLen.int)
     if keyLen > 0:
       if stream.readData(addr key[0], keyLen.int) != keyLen.int: break
 
     if stream.readData(addr valLen, 4) != 4: break
+    if valLen.int > MaxWalRecordField: break
     var value = newSeq[byte](valLen.int)
     if valLen > 0:
       if stream.readData(addr value[0], valLen.int) != valLen.int: break

src/barabadb/storage/wal.nim

@@ -326,8 +326,9 @@ proc rewriteLive*(wal: var WriteAheadLog,
 
   if wal.stream != nil:
     wal.stream.close()
-  if fileExists(wal.path):
-    removeFile(wal.path)
+    wal.stream = nil
+  # Atomic replace: moveFile overwrites the destination on POSIX rename(2).
+  # Do not removeFile first — a crash between unlink and rename would lose the WAL.
   moveFile(tmpPath, wal.path)
   wal.stream = newFileStream(wal.path, fmAppend)
   if wal.stream == nil:
@@ -363,6 +364,7 @@ proc readEntries*(walPath: string, untilTimestamp: uint64 = 0): seq[WalEntry] =
     if s.readData(addr magic, 4) != 4: return
     if s.readData(addr version, 4) != 4: return
     if magic != WALMagic: return
+    const MaxWalRecordField = 64 * 1024 * 1024  # 64 MB
     while not s.atEnd:
       var kind: uint8
       if s.readData(addr kind, 1) != 1: break
@@ -372,11 +374,14 @@ proc readEntries*(walPath: string, untilTimestamp: uint64 = 0): seq[WalEntry] =
         break
       var keyLen: uint32
       if s.readData(addr keyLen, 4) != 4: break
+      if keyLen.int > MaxWalRecordField: break
+      if kind < uint8(wekPut) or kind > uint8(wekCommit): break
       var key = newSeq[byte](keyLen)
       if keyLen > 0:
         if s.readData(addr key[0], int(keyLen)) != int(keyLen): break
       var valLen: uint32
       if s.readData(addr valLen, 4) != 4: break
+      if valLen.int > MaxWalRecordField: break
       var value = newSeq[byte](valLen)
       if valLen > 0:
         if s.readData(addr value[0], int(valLen)) != int(valLen): break

src/baradadb.nim

@@ -23,6 +23,7 @@ import barabadb/core/gossip
 import barabadb/core/replication
 import barabadb/core/disttxn
 import barabadb/core/registry
+import barabadb/core/backup
 import barabadb/tools/repair
 import barabadb/tools/migrate
 
@@ -40,27 +41,18 @@ proc newCompactionManager*(db: LSMTree): CompactionManager =
   result.strategy.rebuildFromLSM(db)
 
 proc applyCompactionResult(db: LSMTree, result: compaction.CompactionResult) =
-  ## Apply compaction output under the caller's lock: update sstables + MANIFEST.
-  ## On Linux, compact may already have unlinked inputs; we still close our mmaps.
+  ## Crash-safe apply: load output while inputs still exist, swap catalog,
+  ## write MANIFEST, then unlink inputs. A crash before MANIFEST leaves the
+  ## old set intact (orphan output is ignored); a crash after MANIFEST leaves
+  ## at worst unlinked-but-closed input files.
   if result.outputTables.len == 0:
     return
 
-  var newSSTables: seq[SSTable] = @[]
-  var removedPaths = initTable[string, bool]()
-  for t in result.inputTables:
-    removedPaths[t.path] = true
-  for sst in db.sstables.mitems:
-    if sst.path notin removedPaths:
-      newSSTables.add(sst)
-    else:
-      # Drop mmap after compact unlinked the path (fd remains valid until close)
-      sst.close()
-
+  var loaded: seq[SSTable] = @[]
   for meta in result.outputTables:
     try:
       var sst = loadSSTable(meta.path)
       let name = splitFile(meta.path).name
-      # Prefer numeric id from filename; otherwise allocate
       let parsed = try: parseInt(name) except CatchableError: -1
       if parsed >= 0:
         sst.id = parsed
@@ -68,11 +60,30 @@ proc applyCompactionResult(db: LSMTree, result: compaction.CompactionResult) =
         sst.id = db.nextSSTableId
         inc db.nextSSTableId
       sst.level = meta.level
-      newSSTables.add(sst)
+      loaded.add(sst)
       db.nextSSTableId = max(db.nextSSTableId, sst.id + 1)
     except CatchableError as e:
       warn("Compaction output SSTable failed to load: " & meta.path & " — " & e.msg)
+      for s in loaded.mitems:
+        s.close()
+      for m in result.outputTables:
+        try: removeFile(m.path) except CatchableError: discard
+      return
 
+  var removedPaths = initTable[string, bool]()
+  for t in result.inputTables:
+    removedPaths[t.path] = true
+
+  var newSSTables: seq[SSTable] = @[]
+  var toDrop: seq[SSTable] = @[]
+  for sst in db.sstables.mitems:
+    if sst.path notin removedPaths:
+      newSSTables.add(sst)
+    else:
+      toDrop.add(sst)
+
+  for sst in loaded:
+    newSSTables.add(sst)
   newSSTables.sort(proc(a, b: SSTable): int = cmp(a.id, b.id))
   db.sstables = newSSTables
   db.needsCompaction = db.countL0() >= L0CompactionTrigger
@@ -83,6 +94,13 @@ proc applyCompactionResult(db: LSMTree, result: compaction.CompactionResult) =
   except CatchableError as e:
     warn("Failed to write MANIFEST after compaction: " & e.msg)
 
+  for sst in toDrop.mitems:
+    try:
+      removeFile(sst.path)
+    except CatchableError as e:
+      warn("Failed to remove compacted SSTable: " & sst.path & " — " & e.msg)
+    sst.close()
+
 proc compact*(cm: CompactionManager) =
   # Gate first (cross-thread), then per-DB write lock
   withStorageGate:
@@ -288,7 +306,7 @@ proc main() =
   # Init structured logger from config
   let logLvl = parseEnum[LogLevel]("ll" & capitalizeAscii(config.logLevel))
   defaultLogger = newLogger(logLvl, config.logFile)
-  info("BaraDB v1.2.0 — Multimodal Database Engine")
+  info("BaraDB v1.3.0 — Multimodal Database Engine")
   info("Storage gate initialized (serializes HTTP/TCP/compaction access)")
 
   # Security check: warn if JWT secret is not configured (non-production only)
@@ -338,6 +356,21 @@ proc main() =
   var raftNet: RaftNetwork = nil
   if config.raftEnabled:
     info("Starting Raft node " & config.raftNodeId & " on port " & $config.raftPort)
+    var raftTls: TLSContext = nil
+    if config.raftTlsEnabled:
+      if config.raftTlsCertFile.len == 0 or config.raftTlsKeyFile.len == 0 or
+         not fileExists(config.raftTlsCertFile) or not fileExists(config.raftTlsKeyFile):
+        raise newException(ValueError,
+          "BARADB_RAFT_TLS_ENABLED=true but cert/key missing " &
+          "(BARADB_RAFT_TLS_CERT_FILE / BARADB_RAFT_TLS_KEY_FILE)")
+      if config.raftTlsVerifyPeer and config.raftTlsCaFile.len > 0 and
+         not fileExists(config.raftTlsCaFile):
+        raise newException(ValueError,
+          "BARADB_RAFT_TLS_VERIFY_PEER=true but CA file missing: " &
+          config.raftTlsCaFile & " (BARADB_RAFT_TLS_CA_FILE)")
+      raftTls = newTLSContext(newTLSConfig(
+        config.raftTlsCertFile, config.raftTlsKeyFile,
+        caFile = config.raftTlsCaFile, verifyPeer = config.raftTlsVerifyPeer))
     let raftDataDir = config.dataDir / "raft"
     createDir(raftDataDir)  # idempotent; loadState reads from it, saveState writes
     # Raft convention: `peers` excludes the node itself (majority math and
@@ -350,6 +383,10 @@ proc main() =
     raftNode.peerAddrs = config.raftPeerAddrs
     if config.raftLogMaxEntries > 0:
       raftNode.logMaxEntries = config.raftLogMaxEntries
+    if config.raftSnapChunkKb > 0:
+      raftNode.snapChunkBytes = config.raftSnapChunkKb * 1024
+    if config.raftPeerStaleMs > 0:
+      raftNode.raftPeerStaleMs = config.raftPeerStaleMs
     tcpServer.raftNode = raftNode  # C3b: executeQuery rejects writes on followers
     httpServer.raftNode = raftNode  # /metrics + /health raft gauges
     # Wire state machine: committed entries update LSM + secondary indexes
@@ -367,13 +404,83 @@ proc main() =
         elif cmd == "ddl":
           applyReplicatedDdl(ctx, cast[string](data))
 
+    # Follower InstallSnapshot restore: swap the default DB's data directory
+    # with the received archive, then reopen it into the same DatabaseInfo
+    # slot (the applyCommand closure above keeps working through the swap).
+    # Runs on the raft async event loop and performs blocking disk I/O
+    # (tar extract + LSM close/reopen); snapshot installs are rare, so we
+    # accept the stall rather than adding a worker round-trip.
+    let defaultDbDir = config.dataDir / "databases" / "default"
+    raftNode.restoreSnapshot = proc(archivePath: string, baseIndex: uint64,
+                                    baseTerm: uint64): bool {.gcsafe.} =
+      # NOTE: core/logging's info/warn are not gcsafe (global logger), so
+      # this callback stays silent; restoreDataDir echoes progress itself.
+      echo "[raft] Installing snapshot (base index ", baseIndex,
+           ", base term ", baseTerm, ")"
+      # gcsafe cast: this runs on the raft event-loop thread (same thread as
+      # the rest of the server); the registry ctxFactory type is not marked
+      # gcsafe, which would otherwise reject the call.
+      {.cast(gcsafe).}:
+        # Hold the storage gate for the whole close/extract/reopen/repoint
+        # sequence: HTTP workers run queries under the same gate, so this
+        # cannot close the LSM out from under an in-flight /query.
+        withStorageGate:
+          try:
+            defaultDbInfo.db.close()
+            # restoreDataDir moves the old dir aside and extracts the archive; on
+            # extraction failure it rolls back automatically. Reopen whatever is
+            # on disk either way so the node is not left with a closed DB.
+            let restored = restoreDataDir(archivePath, defaultDbDir)
+            let reopened = registry.reopenDatabase("default")
+            if reopened:
+              # Serve the (re)opened data. Client connections clone tcpServer.ctx
+              # on accept (cloneForConnection), and reopenDatabase installs a NEW
+              # ctx object in the registry slot — without repointing, queries
+              # keep reading the closed pre-restore LSM (empty results, no
+              # error). The websocket change hook was installed on the previous
+              # ctx object; carry it over.
+              let oldCtx = tcpServer.ctx
+              let newCtx = cast[ExecutionContext](cast[pointer](defaultDbInfo.ctx))
+              newCtx.onChange = oldCtx.onChange
+              tcpServer.db = defaultDbInfo.db
+              tcpServer.ctx = newCtx
+              tcpServer.txnManager = newCtx.txnManager
+              # The HTTP thread reads server.db/ctx only inside
+              # withStorageGate (getRequestDatabaseContext in the /query and
+              # /tables handlers), so repointing them here — while this thread
+              # holds the gate — is race-free against request handlers.
+              httpServer.db = defaultDbInfo.db
+              httpServer.ctx = newCtx
+            result = reopened and restored
+          except CatchableError as e:
+            echo "[raft] Snapshot restore failed: ", e.msg
+            result = false
+
+    # Leader InstallSnapshot send: tar the default DB's data directory into the
+    # path raft picks (dataDir/raft/snap_out_<snapId>.tar). The tar runs here on
+    # the raft event loop under the storage gate; sendSnapshot then gzips it on
+    # a worker thread off the loop (gzipFileAsync) so heartbeats keep flowing.
+    raftNode.buildSnapshot = proc(destPath: string): bool {.gcsafe.} =
+      echo "[raft] Building snapshot tar ", destPath
+      {.cast(gcsafe).}:
+        # Hold the storage gate while tarring the data dir so a concurrent
+        # memtable flush (HTTP /query path) cannot write an SSTable
+        # mid-archive. Compression happens later, off the gate (see
+        # sendSnapshot), so it neither stalls the loop nor blocks applies.
+        withStorageGate:
+          try:
+            result = tarDataDir(defaultDbDir, destPath)
+          except CatchableError as e:
+            echo "[raft] Snapshot build failed: ", e.msg
+            result = false
+
     # Wire RAFT ↔ DistTxn
     wireRaftDistTxn(raftNode, tcpServer)
 
     # Wire replication ↔ DistTxn
     wireReplicationDistTxn(tcpServer.replicationManager, tcpServer.distTxnManager)
 
-    raftNet = newRaftNetwork(raftNode)
+    raftNet = newRaftNetwork(raftNode, raftTls)
     asyncCheck raftNet.run()
 
   # HTTP (hunos) after raft wiring so /metrics can see raftNode

tests/bugfix_test.nim

@@ -4,9 +4,19 @@ import std/os
 import std/tables
 import ../src/barabadb/query/[parser, executor, lexer, ast]
 import ../src/barabadb/query/exec/params
+import ../src/barabadb/query/exec/dml
 import ../src/barabadb/core/types
 import ../src/barabadb/core/config
+import ../src/barabadb/core/replication
+import ../src/barabadb/core/disttxn
+import ../src/barabadb/core/websocket
+import ../src/barabadb/protocol/auth
+import ../src/barabadb/protocol/scram
 import ../src/barabadb/storage/lsm
+import ../src/barabadb/storage/compaction
+import ../src/barabadb/storage/btree
+import std/random
+import std/sets
 
 const testDir = "/tmp/baradb_bugfix_test"
 
@@ -417,6 +427,78 @@ suite "Raft peer address parsing":
       check msg.len > 0
       check bad in msg
 
+suite "Raft put/delete encoding — empty value is not a delete":
+
+  test "PK-only INSERT yields a put pair (deleted == false, empty value)":
+    var ctx = setupCtx()
+    defer: teardown(ctx)
+    discard executeQuery(ctx, parse("CREATE TABLE pkonly (id INTEGER PRIMARY KEY)"))
+    let r = executeQuery(ctx, parse("INSERT INTO pkonly (id) VALUES (1)"))
+    check r.success
+    check r.keyValuePairs.len == 1
+    check r.keyValuePairs[0].value.len == 0
+    check r.keyValuePairs[0].deleted == false
+
+  test "DELETE yields a delete pair (deleted == true)":
+    var ctx = setupCtx()
+    defer: teardown(ctx)
+    discard executeQuery(ctx, parse("INSERT INTO users (id, name) VALUES (1, 'alice')"))
+    let r = executeQuery(ctx, parse("DELETE FROM users WHERE id = 1"))
+    check r.success
+    check r.keyValuePairs.len == 1
+    check r.keyValuePairs[0].deleted == true
+    check r.keyValuePairs[0].value.len == 0
+
+  test "UPDATE yields a put pair (deleted == false)":
+    var ctx = setupCtx()
+    defer: teardown(ctx)
+    discard executeQuery(ctx, parse("INSERT INTO users (id, name) VALUES (1, 'alice')"))
+    let r = executeQuery(ctx, parse("UPDATE users SET name = 'bob' WHERE id = 1"))
+    check r.success
+    check r.keyValuePairs.len == 1
+    check r.keyValuePairs[0].deleted == false
+    check r.keyValuePairs[0].value.len > 0
+
+  test "txn COMMIT pairs carry deleted flag for buffered writes":
+    var ctx = setupCtx()
+    defer: teardown(ctx)
+    discard executeQuery(ctx, parse("CREATE TABLE pkonly (id INTEGER PRIMARY KEY)"))
+    discard executeQuery(ctx, parse("INSERT INTO users (id, name) VALUES (1, 'alice')"))
+    discard executeQuery(ctx, parse("BEGIN"))
+    discard executeQuery(ctx, parse("INSERT INTO pkonly (id) VALUES (7)"))
+    discard executeQuery(ctx, parse("DELETE FROM users WHERE id = 1"))
+    let r = executeQuery(ctx, parse("COMMIT"))
+    check r.success
+    check r.keyValuePairs.len == 2
+    var sawPut = false
+    var sawDelete = false
+    for pair in r.keyValuePairs:
+      if pair.deleted:
+        sawDelete = true
+        check pair.value.len == 0
+      else:
+        sawPut = true
+        check pair.key == "pkonly.id=7"
+        check pair.value.len == 0  # empty value must still be a put
+    check sawPut and sawDelete
+
+  test "apply of a put with empty value keeps the PK-only row":
+    var ctx = setupCtx()
+    defer: teardown(ctx)
+    discard executeQuery(ctx, parse("CREATE TABLE pkonly (id INTEGER PRIMARY KEY)"))
+    let r = executeQuery(ctx, parse("INSERT INTO pkonly (id) VALUES (3)"))
+    check r.success
+    check r.keyValuePairs.len == 1
+    let pair = r.keyValuePairs[0]
+    # Same decode as applyCommand in src/baradadb.nim for a "put" entry.
+    let encoded = pair.key & "\x00" & cast[string](pair.value)
+    let parts = encoded.split("\x00")
+    check parts.len >= 2
+    applyReplicatedPut(ctx, parts[0], cast[seq[byte]](parts[1]))
+    let sel = executeQuery(ctx, parse("SELECT * FROM pkonly WHERE id = 3"))
+    check sel.success
+    check sel.rows.len == 1
+
 suite "Raft write classification":
 
   test "isWrite classifies DML and COMMIT":
@@ -448,3 +530,419 @@ suite "Raft write classification":
     check not isRaftDdl(parse("CREATE DATABASE other").stmts[0])
     check not isRaftDdl(parse("INSERT INTO t (id) VALUES (1)").stmts[0])
     check not isRaftDdl(parse("SELECT 1").stmts[0])
+
+
+suite "Raft TLS config":
+
+  test "default config has raft TLS disabled with empty paths":
+    let cfg = defaultConfig()
+    check cfg.raftTlsEnabled == false
+    check cfg.raftTlsCertFile == ""
+    check cfg.raftTlsKeyFile == ""
+    check cfg.raftTlsCaFile == ""
+    check cfg.raftTlsVerifyPeer == false
+
+  test "env vars parse into raft TLS config":
+    putEnv("BARADB_RAFT_TLS_ENABLED", "true")
+    putEnv("BARADB_RAFT_TLS_CERT_FILE", "/tmp/raft.crt")
+    putEnv("BARADB_RAFT_TLS_KEY_FILE", "/tmp/raft.key")
+    putEnv("BARADB_RAFT_TLS_CA_FILE", "/tmp/raft-ca.crt")
+    putEnv("BARADB_RAFT_TLS_VERIFY_PEER", "1")
+    defer:
+      delEnv("BARADB_RAFT_TLS_ENABLED")
+      delEnv("BARADB_RAFT_TLS_CERT_FILE")
+      delEnv("BARADB_RAFT_TLS_KEY_FILE")
+      delEnv("BARADB_RAFT_TLS_CA_FILE")
+      delEnv("BARADB_RAFT_TLS_VERIFY_PEER")
+    var cfg = defaultConfig()
+    loadConfigFromEnv(cfg)
+    check cfg.raftTlsEnabled == true
+    check cfg.raftTlsCertFile == "/tmp/raft.crt"
+    check cfg.raftTlsKeyFile == "/tmp/raft.key"
+    check cfg.raftTlsCaFile == "/tmp/raft-ca.crt"
+    check cfg.raftTlsVerifyPeer == true
+
+
+suite "Legacy REP payload encoding — empty value is not a delete":
+
+  test "PK-only put (empty value) round-trips as a put, not a delete":
+    ## Regression: the legacy REP receiver used to infer a delete from an empty
+    ## value, so PK-only rows (empty LSM value) vanished on the replica.
+    let decoded = decodeRepPayload(encodeRepPayload(false, "pkonly.id=3", @[]))
+    check decoded.op == ropPut
+    check decoded.key == "pkonly.id=3"
+    check decoded.value.len == 0
+
+  test "delete round-trips as a delete":
+    let decoded = decodeRepPayload(encodeRepPayload(true, "users.id=1", @[]))
+    check decoded.op == ropDelete
+    check decoded.key == "users.id=1"
+    check decoded.value.len == 0
+
+  test "put with a non-empty value preserves the value bytes":
+    let decoded = decodeRepPayload(
+      encodeRepPayload(false, "users.id=1", cast[seq[byte]]("bob")))
+    check decoded.op == ropPut
+    check decoded.key == "users.id=1"
+    check cast[string](decoded.value) == "bob"
+
+  test "value containing a null byte survives the round-trip":
+    ## Decode splits on the FIRST null (the key/value separator) only.
+    let value = @[byte('a'), byte(0), byte('b')]
+    let decoded = decodeRepPayload(encodeRepPayload(false, "k", value))
+    check decoded.op == ropPut
+    check decoded.key == "k"
+    check decoded.value == value
+
+  test "empty or untagged payloads decode as invalid, not delete":
+    check decodeRepPayload(@[]).op == ropInvalid
+    check decodeRepPayload(cast[seq[byte]]("Xfoo")).op == ropInvalid
+
+
+suite "Query operator correctness — audit batch 1":
+
+  test "power operator ** evaluates, not lowered to equality":
+    ## Regression: bkPow used to fall through to `else: irOp = irEq`, so
+    ## `2 ** 3` evaluated as `2 = 3` (false) instead of 8.
+    var ctx = setupCtx()
+    defer: teardown(ctx)
+    discard executeQuery(ctx, parse("INSERT INTO users (id, name) VALUES (1, 'a')"))
+    let r = executeQuery(ctx, parse("SELECT 2 ** 3 AS x FROM users"))
+    check r.success
+    check r.rows.len == 1
+    check parseFloat(valueToString(r.rows[0]["x"])) == 8.0
+
+  test "concat operator ++ concatenates strings":
+    ## Regression: bkConcat also fell through to irEq, so `'a' ++ 'b'`
+    ## evaluated as `'a' = 'b'` (false) instead of "ab".
+    var ctx = setupCtx()
+    defer: teardown(ctx)
+    discard executeQuery(ctx, parse("INSERT INTO users (id, name) VALUES (1, 'a')"))
+    let r = executeQuery(ctx, parse("SELECT 'a' ++ 'b' AS x FROM users"))
+    check r.success
+    check r.rows.len == 1
+    check valueToString(r.rows[0]["x"]) == "ab"
+
+  test "!= is the complement of = for numerically equal values":
+    ## Regression: irNeq short-circuited on string inequality, so `1 != 1.0`
+    ## was true while `1 = 1.0` was also true (not complements).
+    var ctx = setupCtx()
+    defer: teardown(ctx)
+    discard executeQuery(ctx, parse("INSERT INTO users (id, name) VALUES (1, 'alice')"))
+    let eq = executeQuery(ctx, parse("SELECT * FROM users WHERE id = 1.0"))
+    let neq = executeQuery(ctx, parse("SELECT * FROM users WHERE id != 1.0"))
+    check eq.rows.len == 1   # 1 = 1.0 -> true
+    check neq.rows.len == 0  # 1 != 1.0 -> false (old bug returned the row)
+
+
+suite "Query correctness — audit batch 2":
+
+  test "COUNT(DISTINCT) deduplicates values":
+    ## Regression: funcDistinct was parsed but never copied to aggDistinct /
+    ## never consulted during aggregation.
+    var ctx = setupCtx()
+    defer: teardown(ctx)
+    discard executeQuery(ctx, parse("INSERT INTO users (id, name) VALUES (1, 'alice')"))
+    discard executeQuery(ctx, parse("INSERT INTO users (id, name) VALUES (2, 'bob')"))
+    discard executeQuery(ctx, parse("INSERT INTO users (id, name) VALUES (3, 'alice')"))
+    let r = executeQuery(ctx, parse("SELECT COUNT(DISTINCT name) AS c FROM users"))
+    check r.success
+    check r.rows.len == 1
+    check valueToString(r.rows[0]["c"]) == "2"
+
+  test "SUM(DISTINCT) sums unique values only":
+    var ctx = setupCtx()
+    defer: teardown(ctx)
+    discard executeQuery(ctx, parse("INSERT INTO users (id, name) VALUES (1, 'a')"))
+    discard executeQuery(ctx, parse("INSERT INTO users (id, name) VALUES (2, 'b')"))
+    discard executeQuery(ctx, parse("INSERT INTO users (id, name) VALUES (5, 'c')"))
+    # ids 1, 2, 5 — insert another row with id-like values via a number col
+    discard executeQuery(ctx, parse("CREATE TABLE nums (id INTEGER PRIMARY KEY, n INTEGER)"))
+    discard executeQuery(ctx, parse("INSERT INTO nums (id, n) VALUES (1, 10)"))
+    discard executeQuery(ctx, parse("INSERT INTO nums (id, n) VALUES (2, 10)"))
+    discard executeQuery(ctx, parse("INSERT INTO nums (id, n) VALUES (3, 20)"))
+    let r = executeQuery(ctx, parse("SELECT SUM(DISTINCT n) AS s FROM nums"))
+    check r.success
+    check r.rows.len == 1
+    check parseFloat(valueToString(r.rows[0]["s"])) == 30.0
+
+  test "UNION deduplicates without KeyError":
+    ## Regression: set-op dedup used row["$value"] which projected rows lack.
+    var ctx = setupCtx()
+    defer: teardown(ctx)
+    discard executeQuery(ctx, parse("INSERT INTO users (id, name) VALUES (1, 'alice')"))
+    discard executeQuery(ctx, parse("INSERT INTO users (id, name) VALUES (2, 'bob')"))
+    discard executeQuery(ctx, parse("INSERT INTO users (id, name) VALUES (3, 'alice')"))
+    let r = executeQuery(ctx, parse(
+      "SELECT name FROM users WHERE id = 1 UNION SELECT name FROM users WHERE id = 3"))
+    check r.success
+    check r.rows.len == 1
+    check valueToString(r.rows[0]["name"]) == "alice"
+
+  test "INTERSECT returns common rows":
+    var ctx = setupCtx()
+    defer: teardown(ctx)
+    discard executeQuery(ctx, parse("INSERT INTO users (id, name) VALUES (1, 'alice')"))
+    discard executeQuery(ctx, parse("INSERT INTO users (id, name) VALUES (2, 'bob')"))
+    let r = executeQuery(ctx, parse(
+      "SELECT name FROM users WHERE id <= 2 INTERSECT SELECT name FROM users WHERE id = 1"))
+    check r.success
+    check r.rows.len == 1
+    check valueToString(r.rows[0]["name"]) == "alice"
+
+  test "EXCEPT removes right-side rows":
+    var ctx = setupCtx()
+    defer: teardown(ctx)
+    discard executeQuery(ctx, parse("INSERT INTO users (id, name) VALUES (1, 'alice')"))
+    discard executeQuery(ctx, parse("INSERT INTO users (id, name) VALUES (2, 'bob')"))
+    let r = executeQuery(ctx, parse(
+      "SELECT name FROM users EXCEPT SELECT name FROM users WHERE id = 1"))
+    check r.success
+    check r.rows.len == 1
+    check valueToString(r.rows[0]["name"]) == "bob"
+
+  test "MERGE WHEN MATCHED THEN DELETE removes the row":
+    ## Regression: mergeMatchedDelete was parsed but never executed.
+    var ctx = setupCtx()
+    defer: teardown(ctx)
+    discard executeQuery(ctx, parse("CREATE TABLE inv (id INTEGER PRIMARY KEY, qty INTEGER)"))
+    discard executeQuery(ctx, parse("INSERT INTO inv (id, qty) VALUES (1, 10)"))
+    discard executeQuery(ctx, parse("INSERT INTO inv (id, qty) VALUES (2, 20)"))
+    discard executeQuery(ctx, parse("CREATE TABLE deltas (id INTEGER PRIMARY KEY, qty INTEGER)"))
+    discard executeQuery(ctx, parse("INSERT INTO deltas (id, qty) VALUES (1, 0)"))
+    let r = executeQuery(ctx, parse("""
+      MERGE INTO inv AS t
+      USING deltas AS s
+      ON t.id = s.id
+      WHEN MATCHED THEN DELETE
+    """))
+    check r.success
+    check r.affectedRows >= 1
+    let left = executeQuery(ctx, parse("SELECT id FROM inv ORDER BY id"))
+    check left.success
+    check left.rows.len == 1
+    check valueToString(left.rows[0]["id"]) == "2"
+
+  test "semi-sync writeLsn returns 0 when replicas do not ack":
+    var rm = newReplicationManager(rmSemiSync, syncCount = 1)
+    rm.addReplica(newReplica("r1", "10.0.0.1", 9472))
+    rm.connectReplica("r1")
+    let lsn = rm.writeLsn(@[1'u8, 2, 3])
+    check lsn == 0
+
+
+suite "Audit batch 3 — remaining 2026-08 findings":
+
+  test "TLS CA file auto-enables peer verify":
+    putEnv("BARADB_TLS_CA_FILE", "/tmp/ca.crt")
+    defer: delEnv("BARADB_TLS_CA_FILE")
+    var cfg = defaultConfig()
+    loadConfigFromEnv(cfg)
+    check cfg.tlsCaFile == "/tmp/ca.crt"
+    check cfg.tlsVerifyPeer == true
+
+  test "explicit BARADB_TLS_VERIFY_PEER=false wins over CA auto-enable":
+    putEnv("BARADB_TLS_CA_FILE", "/tmp/ca.crt")
+    putEnv("BARADB_TLS_VERIFY_PEER", "false")
+    defer:
+      delEnv("BARADB_TLS_CA_FILE")
+      delEnv("BARADB_TLS_VERIFY_PEER")
+    var cfg = defaultConfig()
+    loadConfigFromEnv(cfg)
+    check cfg.tlsVerifyPeer == false
+
+  test "production TLS without verify is rejected":
+    putEnv("BARADB_ENV", "production")
+    defer: delEnv("BARADB_ENV")
+    var cfg = defaultConfig()
+    cfg.authEnabled = true
+    cfg.jwtSecret = "a".repeat(32)
+    cfg.tlsEnabled = true
+    var msg = ""
+    try:
+      validateProductionConfig(cfg)
+    except ValueError as e:
+      msg = e.msg
+    check "TLS" in msg or "verify" in msg.toLower()
+
+  test "OFFSET without LIMIT returns remaining rows":
+    var ctx = setupCtx()
+    defer: teardown(ctx)
+    discard executeQuery(ctx, parse("INSERT INTO users (id, name) VALUES (1, 'a')"))
+    discard executeQuery(ctx, parse("INSERT INTO users (id, name) VALUES (2, 'b')"))
+    discard executeQuery(ctx, parse("INSERT INTO users (id, name) VALUES (3, 'c')"))
+    let r = executeQuery(ctx, parse("SELECT id FROM users ORDER BY id OFFSET 1"))
+    check r.success
+    check r.rows.len == 2
+    check valueToString(r.rows[0]["id"]) == "2"
+    check valueToString(r.rows[1]["id"]) == "3"
+
+  test "LIMIT 0 returns no rows":
+    var ctx = setupCtx()
+    defer: teardown(ctx)
+    discard executeQuery(ctx, parse("INSERT INTO users (id, name) VALUES (1, 'a')"))
+    let r = executeQuery(ctx, parse("SELECT id FROM users LIMIT 0"))
+    check r.success
+    check r.rows.len == 0
+
+  test "negative LIMIT is clamped to empty, not IndexDefect":
+    var ctx = setupCtx()
+    defer: teardown(ctx)
+    discard executeQuery(ctx, parse("INSERT INTO users (id, name) VALUES (1, 'a')"))
+    let r = executeQuery(ctx, parse("SELECT id FROM users LIMIT -5"))
+    check r.success
+    check r.rows.len == 0
+
+  test "WebSocket decode rejects unmasked client frames":
+    let raw = encodeFrame(0x1, "SUBSCRIBE t", masked = false)
+    let (_, n) = decodeFrame(raw)
+    check n < 0
+
+  test "WebSocket decode accepts masked client frames":
+    let raw = encodeFrame(0x1, "SUBSCRIBE t", masked = true,
+                          maskKey = [1'u8, 2, 3, 4])
+    let (frame, n) = decodeFrame(raw)
+    check n > 0
+    check frame.masked
+    check frame.payload == "SUBSCRIBE t"
+
+  test "WebSocket decode rejects oversized frame length":
+    var raw = newString(10)
+    raw[0] = char(0x81)  # FIN + text
+    raw[1] = char(0xFF)  # MASK + 127
+    # 8-byte length = 2 GiB
+    raw[2] = char(0)
+    raw[3] = char(0)
+    raw[4] = char(0)
+    raw[5] = char(0)
+    raw[6] = char(0x80)
+    raw[7] = char(0)
+    raw[8] = char(0)
+    raw[9] = char(0)
+    let (_, n) = decodeFrame(raw)
+    check n < 0
+
+  test "disttxn prepare against refused port fails closed":
+    var tm = newDistTxnManager()
+    let txn = tm.beginTransaction("coord")
+    txn.addParticipant("n1", "127.0.0.1", 1)
+    check txn.prepare() == false
+    check txn.isAborted
+
+  test "compact leaves input files on disk for catalog apply":
+    let testDir = "/tmp/baradb_bugfix_compact_order"
+    removeDir(testDir)
+    var db = newLSMTree(testDir, 128)
+    defer:
+      db.close()
+      removeDir(testDir)
+    for round in 0 ..< L0CompactionTrigger:
+      db.put("r" & $round, cast[seq[byte]]("v" & $round))
+      db.flush()
+    var cs = newCompactionStrategy(testDir)
+    cs.rebuildFromLSM(db)
+    let cr = cs.compact(0)
+    check cr.outputTables.len == 1
+    check fileExists(cr.outputTables[0].path)
+    for t in cr.inputTables:
+      check fileExists(t.path)
+
+  test "SCRAM unknown user fails without leaking existence":
+    var am = newAuthManager()
+    am.registerScramUser("alice", "wonderland")
+    var msg = ""
+    try:
+      discard am.startScram("n,,n=eve,r=abcnonceabcnonceabcn")
+    except ValueError as e:
+      msg = e.msg
+    check msg == "Authentication failed"
+
+  test "SCRAM rejects mismatched channel binding":
+    var am = newAuthManager()
+    am.registerScramUser("alice", "wonderland")
+    let clientNonce = generateNonce()
+    let clientFirst = "n,,n=alice,r=" & clientNonce
+    let serverFirst = am.startScram(clientFirst)
+    var combinedNonce = ""
+    for part in serverFirst.split(","):
+      if part.startsWith("r="): combinedNonce = part[2..^1]
+    let (ok, err) = am.finishScram("c=AAAA,r=" & combinedNonce & ",p=AA")
+    check ok == false
+    check err == "e=channel-bindings-dont-match"
+
+  test "SCRAM expected cbind for gs2 n is biws":
+    check expectedChannelBinding("n") == "biws"
+
+
+suite "Audit batch 4 — B-tree separator (H10) and NULL equality (L4)":
+
+  test "B-tree separators stay valid after interleaved insert/remove":
+    ## Search routes with `key > sep → right`. After delete, removeRec used to
+    ## copy the right child's first key into the separator (right-min), which
+    ## breaks max(left) <= sep < min(right).
+    var rng = initRand(20260828)
+    var btree = newBTreeIndex[int, string](order = 5)
+    var tracker = initTable[int, seq[string]]()
+    for i in 0..<400:
+      let k = rng.rand(0..80)
+      if rng.rand(0..2) < 2:
+        let v = "v" & $i
+        btree.insert(k, v)
+        if k notin tracker: tracker[k] = @[]
+        tracker[k].add(v)
+      else:
+        if k in tracker and tracker[k].len > 0:
+          let v = tracker[k][0]
+          btree.remove(k, v)
+          tracker[k].del(0)
+          if tracker[k].len == 0: tracker.del(k)
+      if i mod 50 == 49:
+        check btree.checkInvariants().len == 0
+    for k, vals in tracker:
+      check btree.get(k).toHashSet == vals.toHashSet
+
+  test "B-tree sequential fill + prefix delete keeps separators":
+    var btree = newBTreeIndex[int, string](order = 4)
+    for i in 0..<60:
+      btree.insert(i, "v" & $i)
+    for i in 0..<30:
+      btree.remove(i, "v" & $i)
+    check btree.checkInvariants().len == 0
+    for i in 30..<60:
+      check btree.get(i) == @["v" & $i]
+    for i in 0..<30:
+      check btree.get(i).len == 0
+
+  test "NULL = NULL is unknown, not true":
+    var ctx = setupCtx()
+    defer: teardown(ctx)
+    discard executeQuery(ctx, parse("INSERT INTO users (id, name) VALUES (1, NULL)"))
+    discard executeQuery(ctx, parse("INSERT INTO users (id, name) VALUES (2, 'bob')"))
+    let eqNull = executeQuery(ctx, parse("SELECT id FROM users WHERE name = NULL"))
+    check eqNull.success
+    check eqNull.rows.len == 0
+    let isNull = executeQuery(ctx, parse("SELECT id FROM users WHERE name IS NULL"))
+    check isNull.success
+    check isNull.rows.len == 1
+    check valueToString(isNull.rows[0]["id"]) == "1"
+
+  test "NULL != value is unknown so WHERE excludes the row":
+    var ctx = setupCtx()
+    defer: teardown(ctx)
+    discard executeQuery(ctx, parse("INSERT INTO users (id, name) VALUES (1, NULL)"))
+    discard executeQuery(ctx, parse("INSERT INTO users (id, name) VALUES (2, 'bob')"))
+    let r = executeQuery(ctx, parse("SELECT id FROM users WHERE name != 'bob'"))
+    check r.success
+    check r.rows.len == 0
+
+  test "SELECT NULL = NULL yields NULL, IS NULL is true":
+    var ctx = setupCtx()
+    defer: teardown(ctx)
+    discard executeQuery(ctx, parse("INSERT INTO users (id, name) VALUES (1, 'a')"))
+    let eq = executeQuery(ctx, parse("SELECT (NULL = NULL) AS x FROM users"))
+    check eq.success
+    check eq.rows.len == 1
+    check valueToString(eq.rows[0]["x"]) == "\\N"
+    let isn = executeQuery(ctx, parse("SELECT (NULL IS NULL) AS x FROM users"))
+    check isn.success
+    check valueToString(isn.rows[0]["x"]) == "true"

tests/raft_coldnode_e2e_test.nim

@@ -0,0 +1,480 @@
+## Raft cold-node E2E — real 3-node cluster; end-to-end proof of the
+## InstallSnapshot work (compaction, snapshot build/send, follower restore).
+## Starts three actual build/baradadb processes with a tiny raft log
+## (BARADB_RAFT_LOG_MAX_ENTRIES=16) and a short stale window
+## (BARADB_RAFT_PEER_STALE_MS=3000) so the leader compacts quickly past a
+## downed peer's matchIndex.
+##
+## Scenario A: a node is killed, 100 rows are written through the leader
+## (forcing compaction past its matchIndex), the node restarts with its
+## intact data dir and must catch up via InstallSnapshot within 15 s.
+## Scenario B: the node is stopped, its data dir is WIPED, it rejoins with
+## the same node id and must serve the full row set within 20 s.
+##
+## Process-management conventions follow tests/raft_writes_e2e_test.nim
+## (copying is the repo's e2e convention — do not factor out).
+import std/unittest
+import std/osproc
+import std/os
+import std/strtabs
+import std/strutils
+import std/times
+import std/net
+import std/posix
+import std/httpclient
+
+import ../adaptors/nim/baradb_sqlite as sqlite
+
+const
+  BinaryPath = "./build/baradadb"
+  LeaderMarker = "became leader"
+  SnapshotMarker = "Installing snapshot"
+  TableName = "cold_test"
+  RowCount = 100
+
+type
+  NodeProc = object
+    id: string
+    clientPort: int
+    raftPort: int
+    peers: string
+    clientPeers: string
+    p: Process
+    dataDir: string
+    output: string
+    alive: bool
+
+proc drainOutput(n: var NodeProc) =
+  ## Reads whatever the child has written so far. The pipe was set O_NONBLOCK
+  ## at start, so this never blocks — a hung read is impossible here.
+  var tmp: array[8192, char]
+  while true:
+    let count = posix.read(n.p.outputHandle.cint, tmp[0].addr, tmp.len)
+    if count <= 0: break
+    for i in 0 ..< count: n.output.add tmp[i]
+
+proc drainAll(nodes: var seq[NodeProc]) =
+  for n in nodes.mitems:
+    if n.p != nil: n.drainOutput()
+
+proc dumpAll(nodes: var seq[NodeProc]) =
+  ## Debuggability: on failure, everything the nodes said.
+  nodes.drainAll()
+  for n in nodes:
+    echo "===== output of ", n.id, " (port ", n.clientPort, ") ====="
+    echo n.output
+
+proc portOpen(port: int): bool =
+  var s: Socket
+  try:
+    s = newSocket()
+    s.connect("127.0.0.1", Port(port), timeout = 250)
+    s.close()
+    result = true
+  except CatchableError:
+    if s != nil: s.close()
+    result = false
+
+proc killNode(n: var NodeProc) =
+  if n.p != nil and n.alive:
+    try:
+      n.p.terminate()
+      discard n.p.waitForExit()
+    except CatchableError:
+      discard
+    n.alive = false
+
+proc leaderTerms(output: string): seq[int] =
+  ## All terms this node logged leadership for ("became leader for term T").
+  var pos = 0
+  while true:
+    let idx = output.find(LeaderMarker, pos)
+    if idx < 0: break
+    let tIdx = output.find("term ", idx)
+    if tIdx < 0: break
+    let numStart = tIdx + 5
+    var numEnd = numStart
+    while numEnd < output.len and output[numEnd] in Digits: inc numEnd
+    if numEnd > numStart:
+      result.add(parseInt(output[numStart ..< numEnd]))
+    pos = numEnd
+
+proc maxLeader(nodes: seq[NodeProc]): tuple[idx, term: int] =
+  ## Node that logged leadership for the highest term seen so far.
+  result = (-1, 0)
+  for i in 0 ..< nodes.len:
+    for t in leaderTerms(nodes[i].output):
+      if t > result.term: result = (i, t)
+
+proc drainFor(nodes: var seq[NodeProc], ms: int) =
+  let start = getTime()
+  while getTime() - start < initDuration(milliseconds = ms):
+    nodes.drainAll()
+    sleep(50)
+
+proc openClient(port: int): DbConn =
+  ## Connect with retries — the port may accept TCP before the DB is usable.
+  for i in 0 ..< 50:
+    try:
+      return open("127.0.0.1:" & $port, "", "", "default")
+    except CatchableError:
+      sleep(100)
+  raise newException(IOError, "cannot connect to port " & $port)
+
+proc rowCountOn(port: int): int =
+  ## SELECT COUNT(*) — raises while the table is not there yet (e.g. before
+  ## the snapshot restore has landed); callers poll and tolerate that.
+  let db = openClient(port)
+  defer: db.close()
+  parseInt(db.getValue(sql("SELECT COUNT(*) FROM " & TableName)))
+
+proc fetchMetric(port: int, name: string): int =
+  ## GET /metrics on the node's HTTP port (clientPort + 440) and parse
+  ## `name{...} <value>`. Returns -1 on any error or when absent.
+  let client = newHttpClient(timeout = 1500)
+  defer: client.close()
+  try:
+    let body = client.getContent("http://127.0.0.1:" & $(port + 440) & "/metrics")
+    for line in body.splitLines():
+      if line.startsWith(name & "{"):
+        let parts = line.splitWhitespace()
+        if parts.len >= 2:
+          return parseInt(parts[^1])
+  except CatchableError:
+    discard
+  return -1
+
+proc startNode(id, dataDir: string, clientPort, raftPort: int,
+               peers, clientPeers: string): NodeProc =
+  ## Boots one build/baradadb process. Takes the data dir explicitly so a
+  ## node can be restarted with the same dir (scenario A) or a wiped dir
+  ## (scenario B). Compaction/stale-window env goes on EVERY node.
+  createDir(dataDir)
+  var env = newStringTable()
+  for key, val in envPairs():
+    env[key] = val
+  env["BARADB_PORT"] = $clientPort
+  env["BARADB_RAFT_ENABLED"] = "true"
+  env["BARADB_RAFT_PORT"] = $raftPort
+  env["BARADB_RAFT_NODE_ID"] = id
+  env["BARADB_RAFT_PEERS"] = peers
+  env["BARADB_RAFT_CLIENT_PEERS"] = clientPeers
+  env["BARADB_RAFT_LOG_MAX_ENTRIES"] = "16"
+  env["BARADB_RAFT_PEER_STALE_MS"] = "3000"
+  env["BARADB_DATA_DIR"] = dataDir
+  env["BARADB_LOG_LEVEL"] = "info"
+  let p = startProcess(BinaryPath, env = env,
+                       options = {poStdErrToStdOut, poDaemon})
+  discard fcntl(p.outputHandle.cint, F_SETFL,
+                fcntl(p.outputHandle.cint, F_GETFL) or O_NONBLOCK)
+  NodeProc(id: id, clientPort: clientPort, raftPort: raftPort,
+           peers: peers, clientPeers: clientPeers, p: p,
+           dataDir: dataDir, alive: true)
+
+proc restartNode(n: var NodeProc, wipe: bool) =
+  ## Kill, then boot the same node id again. wipe=false keeps the data dir
+  ## (cold-node return); wipe=true deletes it first (fresh node rejoin).
+  n.killNode()
+  if n.p != nil: n.p.close()
+  if wipe:
+    removeDir(n.dataDir)
+  let fresh = startNode(n.id, n.dataDir, n.clientPort, n.raftPort,
+                        n.peers, n.clientPeers)
+  n.p = fresh.p
+  n.alive = true
+
+proc waitReady(n: NodeProc, deadlineSec: int): bool =
+  let start = getTime()
+  while getTime() - start < initDuration(seconds = deadlineSec):
+    if portOpen(n.clientPort):
+      return true
+    sleep(100)
+  return false
+
+proc runColdNodeScenario() =
+  ## Fatal phase failures dump all captured node output, record a test
+  ## failure, and return; cleanup happens in the finally below either way.
+  let tstamp = getTime().toUnix.int
+  # Port base per the T11 brief: distinct from raft_writes_e2e_test
+  # (46000+mod4000) and raft_tls_e2e_test (54000+mod4000). Client ports are
+  # spaced by 10 because the server derives HTTP (port+440), WS (port+441)
+  # and gossip (raftPort+100) ports — consecutive client ports collide.
+  let cbase = 58000 + (tstamp mod 4000)
+  let rbase = cbase + 100
+  let peers = "n1@127.0.0.1:" & $(rbase + 1) &
+              ",n2@127.0.0.1:" & $(rbase + 2) &
+              ",n3@127.0.0.1:" & $(rbase + 3)
+  # SQL client ports for transparent leader write forwarding.
+  let clientPeers = "n1@127.0.0.1:" & $(cbase + 10) &
+                    ",n2@127.0.0.1:" & $(cbase + 20) &
+                    ",n3@127.0.0.1:" & $(cbase + 30)
+
+  var nodes: seq[NodeProc]
+
+  try:
+    # Node starts live inside the try so a raise from start #2/#3 still
+    # reaches the cleanup in the finally below.
+    for i in 1 .. 3:
+      let id = "n" & $i
+      let dataDir = getTempDir() / "baradb_raft_coldnode_e2e_" & $tstamp & "_" & id
+      nodes.add startNode(id, dataDir, cbase + i * 10, rbase + i,
+                          peers, clientPeers)
+
+    # Readiness: all three client ports accept TCP connections (10s each).
+    for i in 0 ..< nodes.len:
+      if not waitReady(nodes[i], 10):
+        echo "node ", nodes[i].id, " never became ready"
+        dumpAll(nodes)
+        fail()
+        return
+
+    # Election: timeouts are 150-300ms, heartbeat 50ms — a leader should
+    # emerge within ~2s; 10s deadline for margin.
+    var elected = false
+    let electStart = getTime()
+    while getTime() - electStart < initDuration(seconds = 10):
+      nodes.drainAll()
+      if maxLeader(nodes).idx >= 0:
+        elected = true
+        break
+      sleep(50)
+    if not elected:
+      echo "no leader elected within 10s"
+      dumpAll(nodes)
+      fail()
+      return
+
+    # Settle and require stability, same as raft_writes_e2e_test.
+    nodes.drainFor(2000)
+    let (leaderIdx, leaderTerm) = maxLeader(nodes)
+    nodes.drainFor(1000)
+    let (stableIdx, stableTerm) = maxLeader(nodes)
+    if stableIdx != leaderIdx or stableTerm != leaderTerm:
+      echo "cluster unstable: leadership moved from ", nodes[leaderIdx].id,
+           " (term ", leaderTerm, ") to ", nodes[stableIdx].id,
+           " (term ", stableTerm, ")"
+      dumpAll(nodes)
+      fail()
+      return
+    echo "leader elected: ", nodes[leaderIdx].id, " (term ", leaderTerm, ")"
+
+    # The cold node: n3 unless n3 is the leader (then n1) — it must be a
+    # follower so killing it never forces a re-election.
+    let coldIdx = (if leaderIdx == 2: 0 else: 2)
+    let otherIdx = 3 - leaderIdx - coldIdx
+    echo "cold node: ", nodes[coldIdx].id, "; surviving follower: ",
+         nodes[otherIdx].id
+
+    # Schema: CREATE TABLE goes through the raft "ddl" log (C3c). Create on
+    # the leader; wait until the cold node has applied it before killing it.
+    block:
+      let db = openClient(nodes[leaderIdx].clientPort)
+      try:
+        db.exec(sql("CREATE TABLE " & TableName & " (id INT PRIMARY KEY, name STRING)"))
+      except CatchableError as e:
+        echo "leader CREATE TABLE failed: ", e.msg
+        dumpAll(nodes)
+        fail()
+        return
+      db.close()
+    block:
+      let start = getTime()
+      var ready = false
+      while getTime() - start < initDuration(seconds = 5):
+        try:
+          discard rowCountOn(nodes[coldIdx].clientPort)
+          ready = true
+          break
+        except CatchableError:
+          sleep(100)
+      if not ready:
+        echo "cold node ", nodes[coldIdx].id,
+             " never applied CREATE TABLE within 5s"
+        dumpAll(nodes)
+        fail()
+        return
+    echo "schema replicated to cold node ", nodes[coldIdx].id
+
+    # ---- Kill the cold node; write 100 rows through the leader. ----
+    # With logMaxEntries=16 and peerStaleMs=3000 the leader compacts past the
+    # dead node's matchIndex partway through the writes, so the node can only
+    # catch up via InstallSnapshot on return.
+    killNode(nodes[coldIdx])
+    echo "cold node ", nodes[coldIdx].id, " killed; writing ", RowCount, " rows"
+    block:
+      let db = openClient(nodes[leaderIdx].clientPort)
+      var atRow = 0
+      try:
+        for i in 1 .. RowCount:
+          atRow = i
+          db.exec(sql("INSERT INTO " & TableName & " (id, name) VALUES (" &
+                      $i & ", 'row-" & $i & "')"))
+      except CatchableError as e:
+        echo "leader INSERT failed at row ", atRow, ": ", e.msg
+        dumpAll(nodes)
+        fail()
+        return
+      db.close()
+    echo RowCount, " rows committed with ", nodes[coldIdx].id, " down"
+
+    # Compaction evidence on the leader: the in-memory log stayed bounded
+    # (<= 64, far below the 100 entries written) and the snapshot base moved.
+    nodes.drainAll()
+    let leaderLogLen = fetchMetric(nodes[leaderIdx].clientPort,
+                                   "baradb_raft_log_entries")
+    let leaderSnapIdx = fetchMetric(nodes[leaderIdx].clientPort,
+                                    "baradb_raft_snapshot_index")
+    echo "leader after writes: log_entries=", leaderLogLen,
+         " snapshot_index=", leaderSnapIdx
+    if leaderLogLen < 0 or leaderLogLen > 64:
+      echo "leader raft log not bounded: baradb_raft_log_entries=",
+           leaderLogLen, " (want <= 64)"
+      dumpAll(nodes)
+      fail()
+      return
+    if leaderSnapIdx <= 0:
+      echo "leader never compacted: baradb_raft_snapshot_index=",
+           leaderSnapIdx, " (want > 0)"
+      dumpAll(nodes)
+      fail()
+      return
+    echo "leader log stayed bounded and compaction advanced"
+
+    # Sanity: leader and surviving follower agree on the row count.
+    let wantCount = rowCountOn(nodes[leaderIdx].clientPort)
+    if wantCount != RowCount:
+      echo "leader count=", wantCount, " want ", RowCount
+      dumpAll(nodes)
+      fail()
+      return
+
+    # ---- Scenario A: cold node returns with its intact data dir. ----
+    nodes[coldIdx].output.setLen(0)  # fresh log for the snapshot marker scan
+    restartNode(nodes[coldIdx], wipe = false)
+    if not waitReady(nodes[coldIdx], 10):
+      echo "cold node ", nodes[coldIdx].id, " never became ready after restart"
+      dumpAll(nodes)
+      fail()
+      return
+    echo "scenario A: ", nodes[coldIdx].id, " restarted with intact data dir"
+
+    # Within 15s: snapshot evidence (metric on the cold node, or its restore
+    # log line) AND data convergence with the leader.
+    block:
+      let start = getTime()
+      var snapSeen = false
+      var converged = false
+      while getTime() - start < initDuration(seconds = 15):
+        nodes.drainAll()
+        if not snapSeen:
+          snapSeen = fetchMetric(nodes[coldIdx].clientPort,
+                                 "baradb_raft_snapshot_index") > 0 or
+                     nodes[coldIdx].output.contains(SnapshotMarker)
+        if not converged:
+          try:
+            converged = rowCountOn(nodes[coldIdx].clientPort) == wantCount
+          except CatchableError:
+            discard
+        if snapSeen and converged: break
+        sleep(200)
+      if not snapSeen:
+        echo "scenario A: no snapshot evidence on ", nodes[coldIdx].id,
+             " (baradb_raft_snapshot_index stayed 0, no restore log line)"
+        dumpAll(nodes)
+        fail()
+        return
+      if not converged:
+        echo "scenario A: ", nodes[coldIdx].id,
+             " count did not converge to ", wantCount, " within 15s"
+        dumpAll(nodes)
+        fail()
+        return
+    echo "scenario A: snapshot installed, count converged to ", wantCount
+
+    # Spot-check a few ids survived the snapshot round-trip.
+    block:
+      let db = openClient(nodes[coldIdx].clientPort)
+      defer: db.close()
+      for i in [1, 42, RowCount]:
+        let v = db.getValue(sql("SELECT name FROM " & TableName &
+                                " WHERE id = " & $i))
+        if v != "row-" & $i:
+          echo "scenario A: spot-check id=", i, " got '", v,
+               "' want 'row-", i, "'"
+          dumpAll(nodes)
+          fail()
+          return
+    echo "scenario A: spot-checks passed"
+
+    # ---- Scenario B: wiped node rejoins with the same node id. ----
+    nodes[coldIdx].output.setLen(0)
+    restartNode(nodes[coldIdx], wipe = true)
+    if not waitReady(nodes[coldIdx], 10):
+      echo "wiped node ", nodes[coldIdx].id, " never became ready after restart"
+      dumpAll(nodes)
+      fail()
+      return
+    echo "scenario B: ", nodes[coldIdx].id, " restarted with a wiped data dir"
+
+    # Within 20s: snapshot → catch-up, then the full row set.
+    block:
+      let start = getTime()
+      var snapSeen = false
+      var converged = false
+      while getTime() - start < initDuration(seconds = 20):
+        nodes.drainAll()
+        if not snapSeen:
+          snapSeen = fetchMetric(nodes[coldIdx].clientPort,
+                                 "baradb_raft_snapshot_index") > 0 or
+                     nodes[coldIdx].output.contains(SnapshotMarker)
+        if not converged:
+          try:
+            converged = rowCountOn(nodes[coldIdx].clientPort) == wantCount
+          except CatchableError:
+            discard
+        if snapSeen and converged: break
+        sleep(200)
+      if not snapSeen:
+        echo "scenario B: no snapshot evidence on wiped ", nodes[coldIdx].id
+        dumpAll(nodes)
+        fail()
+        return
+      if not converged:
+        echo "scenario B: wiped ", nodes[coldIdx].id,
+             " count did not converge to ", wantCount, " within 20s"
+        dumpAll(nodes)
+        fail()
+        return
+    echo "scenario B: wiped node converged to ", wantCount, " rows"
+
+    # Spot-check again on the wiped node.
+    block:
+      let db = openClient(nodes[coldIdx].clientPort)
+      defer: db.close()
+      for i in [1, 42, RowCount]:
+        let v = db.getValue(sql("SELECT name FROM " & TableName &
+                                " WHERE id = " & $i))
+        if v != "row-" & $i:
+          echo "scenario B: spot-check id=", i, " got '", v,
+               "' want 'row-", i, "'"
+          dumpAll(nodes)
+          fail()
+          return
+    echo "scenario B: spot-checks passed"
+  finally:
+    for n in nodes.mitems:
+      n.killNode()
+      if n.p != nil: n.p.close()
+      removeDir(n.dataDir)
+
+suite "Raft cold-node E2E":
+  test "compacted-away node and wiped node rejoin and converge":
+    if not fileExists(BinaryPath):
+      if getEnv("CI").len > 0:
+        echo "[FAIL] ", BinaryPath, " missing under CI — build step broken?"
+        fail()
+      else:
+        echo "[SKIP] ", BinaryPath, " missing — run `nimble test` (builds the server first)"
+        skip()
+    else:
+      runColdNodeScenario()

tests/raft_e2e_test.nim

@@ -208,7 +208,11 @@ proc runClusterScenario() =
 suite "Raft E2E cluster":
   test "3-node election and failover":
     if not fileExists(BinaryPath):
-      echo "[SKIP] ", BinaryPath, " missing — run `nimble test` (builds the server first)"
-      skip()
+      if getEnv("CI").len > 0:
+        echo "[FAIL] ", BinaryPath, " missing under CI — build step broken?"
+        fail()
+      else:
+        echo "[SKIP] ", BinaryPath, " missing — run `nimble test` (builds the server first)"
+        skip()
     else:
       runClusterScenario()

tests/raft_failover_load_e2e_test.nim

@@ -0,0 +1,435 @@
+## Raft failover-under-load E2E — real 3-node cluster over the TCP transport.
+## A writer thread hammers the current leader with INSERTs; once 50 writes
+## have been acknowledged the leader is killed mid-load. Asserts:
+##   A (availability): a survivor accepts an INSERT within 10s of the kill.
+##   B (durability):   every acknowledged id is present on BOTH survivors
+##                     once the new leader is stable and the remaining
+##                     follower has caught up.
+## Process-management conventions follow tests/raft_writes_e2e_test.nim;
+## client access follows tests/nimforum_smoke_test.nim.
+import std/unittest
+import std/osproc
+import std/os
+import std/strtabs
+import std/strutils
+import std/sequtils
+import std/times
+import std/net
+import std/posix
+import std/sets
+import std/locks
+import std/typedthreads
+
+import ../adaptors/nim/baradb_sqlite as sqlite
+
+const
+  BinaryPath = "./build/baradadb"
+  LeaderMarker = "became leader"
+  AckTarget = 50        # kill the leader once this many writes are acked
+  ProbeId = 1000001     # availability-probe row id (clear of writer's 1..N)
+
+type
+  NodeProc = object
+    id: string
+    clientPort: int
+    p: Process
+    dataDir: string
+    output: string
+    alive: bool
+
+# Shared writer-thread state, passed by pointer into the thread (same
+# convention as tests/test_storage_hardening.nim); all access goes through
+# the lock.
+type
+  WriterArgs = object
+    lock: ptr Lock
+    acked: ptr seq[int]  # ids whose INSERT was acknowledged by the cluster
+    stop: ptr bool       # main thread sets this to end the writer loop
+    ports: array[3, int] # candidate client ports (leader first, then survivors)
+
+proc drainOutput(n: var NodeProc) =
+  ## Reads whatever the child has written so far. The pipe was set O_NONBLOCK
+  ## at start, so this never blocks — a hung read is impossible here.
+  var tmp: array[8192, char]
+  while true:
+    let count = posix.read(n.p.outputHandle.cint, tmp[0].addr, tmp.len)
+    if count <= 0: break
+    for i in 0 ..< count: n.output.add tmp[i]
+
+proc drainAll(nodes: var seq[NodeProc]) =
+  for n in nodes.mitems:
+    if n.p != nil: n.drainOutput()
+
+proc dumpAll(nodes: var seq[NodeProc]) =
+  ## Debuggability: on failure, everything the nodes said.
+  nodes.drainAll()
+  for n in nodes:
+    echo "===== output of ", n.id, " (port ", n.clientPort, ") ====="
+    echo n.output
+
+proc portOpen(port: int): bool =
+  var s: Socket
+  try:
+    s = newSocket()
+    s.connect("127.0.0.1", Port(port), timeout = 250)
+    s.close()
+    result = true
+  except CatchableError:
+    if s != nil: s.close()
+    result = false
+
+proc killNode(n: var NodeProc) =
+  if n.p != nil and n.alive:
+    try:
+      n.p.terminate()
+      discard n.p.waitForExit()
+    except CatchableError:
+      discard
+    n.alive = false
+
+proc leaderTerms(output: string): seq[int] =
+  ## All terms this node logged leadership for ("became leader for term T").
+  var pos = 0
+  while true:
+    let idx = output.find(LeaderMarker, pos)
+    if idx < 0: break
+    let tIdx = output.find("term ", idx)
+    if tIdx < 0: break
+    let numStart = tIdx + 5
+    var numEnd = numStart
+    while numEnd < output.len and output[numEnd] in Digits: inc numEnd
+    if numEnd > numStart:
+      result.add(parseInt(output[numStart ..< numEnd]))
+    pos = numEnd
+
+proc maxLeader(nodes: seq[NodeProc]): tuple[idx, term: int] =
+  ## Node that logged leadership for the highest term seen so far.
+  result = (-1, 0)
+  for i in 0 ..< nodes.len:
+    for t in leaderTerms(nodes[i].output):
+      if t > result.term: result = (i, t)
+
+proc drainFor(nodes: var seq[NodeProc], ms: int) =
+  let start = getTime()
+  while getTime() - start < initDuration(milliseconds = ms):
+    nodes.drainAll()
+    sleep(50)
+
+proc openClient(port: int): DbConn =
+  ## Connect with retries — the port may accept TCP before the DB is usable.
+  for i in 0 ..< 50:
+    try:
+      return open("127.0.0.1:" & $port, "", "", "default")
+    except CatchableError:
+      sleep(100)
+  raise newException(IOError, "cannot connect to port " & $port)
+
+proc countRows(port: int): int =
+  ## Row count of load_test on `port`, or -1 on any error (not ready yet).
+  result = -1
+  try:
+    let db = openClient(port)
+    try:
+      result = parseInt(db.getValue(sql"SELECT count(*) FROM load_test"))
+    finally:
+      db.close()
+  except CatchableError:
+    discard
+
+proc allIds(port: int): HashSet[int] =
+  ## Every id in load_test on `port`. Raises on error.
+  let db = openClient(port)
+  try:
+    for row in db.getAllRows(sql"SELECT id FROM load_test"):
+      if row.len >= 1 and row[0].len > 0:
+        result.incl(parseInt(row[0]))
+  finally:
+    db.close()
+
+proc writerLoop(args: WriterArgs) {.thread.} =
+  ## INSERTs n = 1, 2, ... against the cluster. Every acknowledged n is
+  ## appended to args.acked. On any error the client is dropped and reopened
+  ## against the next candidate port — the documented client retry contract.
+  ## Writes against followers are forwarded to the leader by the server.
+  var n = 1
+  var portIdx = 0
+  var db: DbConn
+  while true:
+    withLock args.lock[]:
+      if args.stop[]: break
+    if cast[pointer](db) == nil:
+      let port = args.ports[portIdx]
+      try:
+        db = open("127.0.0.1:" & $port, "", "", "default")
+      except CatchableError:
+        portIdx = (portIdx + 1) mod args.ports.len
+        sleep(50)
+        continue
+    try:
+      db.exec(sql("INSERT INTO load_test (id) VALUES (" & $n & ")"))
+      withLock args.lock[]:
+        args.acked[].add n
+      inc n
+    except CatchableError:
+      if cast[pointer](db) != nil:
+        try: db.close()
+        except CatchableError: discard
+      db = default(DbConn)
+      portIdx = (portIdx + 1) mod args.ports.len
+      sleep(50)
+  if cast[pointer](db) != nil:
+    try: db.close()
+    except CatchableError: discard
+
+proc runFailoverLoadScenario() =
+  ## Fatal phase failures dump all captured node output, record a test
+  ## failure, and return; cleanup happens in the finally below either way.
+  let tstamp = getTime().toUnix.int
+  # Port bases: distinct from nimforum_smoke_test (35000+mod10000),
+  # raft_e2e_test (41000+mod5000) and raft_writes_e2e_test (46000+mod4000).
+  # Client ports are spaced by 10 because the server derives HTTP (port+440),
+  # WS (port+441) and gossip (raftPort+100) ports — consecutive client ports
+  # collide.
+  let cbase = 50000 + (tstamp mod 4000)
+  let rbase = cbase + 100
+  let peers = "n1@127.0.0.1:" & $(rbase + 1) &
+              ",n2@127.0.0.1:" & $(rbase + 2) &
+              ",n3@127.0.0.1:" & $(rbase + 3)
+  # SQL client ports for transparent leader write forwarding.
+  let clientPeers = "n1@127.0.0.1:" & $(cbase + 10) &
+                    ",n2@127.0.0.1:" & $(cbase + 20) &
+                    ",n3@127.0.0.1:" & $(cbase + 30)
+
+  var nodes: seq[NodeProc]
+  for i in 1 .. 3:
+    let id = "n" & $i
+    let dataDir = getTempDir() / "baradb_raft_failover_load_e2e_" & $tstamp & "_" & id
+    createDir(dataDir)
+    var env = newStringTable()
+    for key, val in envPairs():
+      env[key] = val
+    env["BARADB_PORT"] = $(cbase + i * 10)
+    env["BARADB_RAFT_ENABLED"] = "true"
+    env["BARADB_RAFT_PORT"] = $(rbase + i)
+    env["BARADB_RAFT_NODE_ID"] = id
+    env["BARADB_RAFT_PEERS"] = peers
+    env["BARADB_RAFT_CLIENT_PEERS"] = clientPeers
+    env["BARADB_DATA_DIR"] = dataDir
+    env["BARADB_LOG_LEVEL"] = "info"
+    let p = startProcess(BinaryPath, env = env,
+                         options = {poStdErrToStdOut, poDaemon})
+    discard fcntl(p.outputHandle.cint, F_SETFL,
+                  fcntl(p.outputHandle.cint, F_GETFL) or O_NONBLOCK)
+    nodes.add NodeProc(id: id, clientPort: cbase + i * 10, p: p,
+                       dataDir: dataDir, alive: true)
+
+  var
+    writer: Thread[WriterArgs]
+    writerStarted = false
+    wLock: Lock
+    wAcked: seq[int]
+    wStop = false
+  initLock(wLock)
+  try:
+    # Readiness: all three client ports accept TCP connections (10s each).
+    for i in 0 ..< nodes.len:
+      let readyStart = getTime()
+      var ok = false
+      while getTime() - readyStart < initDuration(seconds = 10):
+        if portOpen(nodes[i].clientPort):
+          ok = true
+          break
+        sleep(100)
+      if not ok:
+        echo "node ", nodes[i].id, " never became ready"
+        dumpAll(nodes)
+        fail()
+        return
+
+    # Election: timeouts are 150-300ms, heartbeat 50ms — a leader should
+    # emerge within ~2s; 10s deadline for margin.
+    var elected = false
+    let electStart = getTime()
+    while getTime() - electStart < initDuration(seconds = 10):
+      nodes.drainAll()
+      if maxLeader(nodes).idx >= 0:
+        elected = true
+        break
+      sleep(50)
+    if not elected:
+      echo "no leader elected within 10s"
+      dumpAll(nodes)
+      fail()
+      return
+
+    # Settle and require stability, same as raft_writes_e2e_test.
+    nodes.drainFor(2000)
+    let (leaderIdx, leaderTerm) = maxLeader(nodes)
+    nodes.drainFor(1000)
+    let (stableIdx, stableTerm) = maxLeader(nodes)
+    if stableIdx != leaderIdx or stableTerm != leaderTerm:
+      echo "cluster unstable: leadership moved from ", nodes[leaderIdx].id,
+           " (term ", leaderTerm, ") to ", nodes[stableIdx].id,
+           " (term ", stableTerm, ")"
+      dumpAll(nodes)
+      fail()
+      return
+    echo "leader elected: ", nodes[leaderIdx].id, " (term ", leaderTerm, ")"
+
+    # Schema goes through the raft ddl log on the leader (C3c).
+    block:
+      let db = openClient(nodes[leaderIdx].clientPort)
+      try:
+        db.exec(sql"CREATE TABLE load_test (id INT PRIMARY KEY)")
+      except CatchableError as e:
+        echo "leader CREATE TABLE failed: ", e.msg
+        dumpAll(nodes)
+        fail()
+        return
+      db.close()
+    echo "leader schema committed via raft ddl"
+
+    # Load phase: sustained INSERTs from a background writer thread.
+    let wArgs = WriterArgs(
+      lock: addr wLock, acked: addr wAcked, stop: addr wStop,
+      ports: [nodes[leaderIdx].clientPort,
+              nodes[(leaderIdx + 1) mod 3].clientPort,
+              nodes[(leaderIdx + 2) mod 3].clientPort])
+    createThread(writer, writerLoop, wArgs)
+    writerStarted = true
+
+    # Wait until AckTarget writes are acknowledged (30s deadline).
+    var ackedAtKill = 0
+    let loadStart = getTime()
+    while getTime() - loadStart < initDuration(seconds = 30):
+      withLock wLock:
+        ackedAtKill = wAcked.len
+      if ackedAtKill >= AckTarget: break
+      sleep(20)
+    if ackedAtKill < AckTarget:
+      echo "only ", ackedAtKill, " writes acked within 30s (need ", AckTarget, ")"
+      dumpAll(nodes)
+      fail()
+      return
+    echo "load phase: ", ackedAtKill, " writes acknowledged, killing leader ",
+         nodes[leaderIdx].id
+
+    # Kill the leader mid-load.
+    killNode(nodes[leaderIdx])
+    let killTime = getTime()
+
+    # Assert A (availability): a survivor accepts an INSERT within 10s of
+    # the kill. Probe both survivors; keep the one that answers. Each attempt
+    # uses a fresh id: an attempt may commit but lose its response during the
+    # failover, and retrying the same id would then loop on UNIQUE violations.
+    var writerSurvivor = -1
+    var writeErr = ""
+    var probeAttempt = 0
+    while getTime() - killTime < initDuration(seconds = 10):
+      for i in 0 ..< nodes.len:
+        if i == leaderIdx: continue
+        inc probeAttempt
+        let probeId = ProbeId + probeAttempt
+        try:
+          let db = openClient(nodes[i].clientPort)
+          try:
+            db.exec(sql("INSERT INTO load_test (id) VALUES (" & $probeId & ")"))
+            writerSurvivor = i
+          finally:
+            db.close()
+          if writerSurvivor >= 0: break
+        except CatchableError as e:
+          writeErr = e.msg
+          # "not leader" / commit timeout / connection blips — keep probing.
+      if writerSurvivor >= 0: break
+      sleep(100)
+    if writerSurvivor < 0:
+      echo "ASSERT A FAILED: no survivor accepted a write within 10s of the kill",
+           (if writeErr.len > 0: " (last error: " & writeErr & ")" else: "")
+      dumpAll(nodes)
+      fail()
+      return
+    let availMs = inMilliseconds(getTime() - killTime)
+    echo "availability: ", nodes[writerSurvivor].id,
+         " accepted a write ", availMs, "ms after the kill"
+
+    # Stop the writer thread and snapshot what was acknowledged.
+    withLock wLock:
+      wStop = true
+    joinThreads(writer)
+    writerStarted = false
+    withLock wLock:
+      ackedAtKill = wAcked.len
+    echo "writer stopped; ", ackedAtKill, " total acknowledged writes"
+
+    # Let the new leader stabilize and the remaining follower catch up:
+    # poll until both survivors agree on a row count that covers every
+    # acknowledged write (10s deadline). The >= ackedAtKill guard prevents
+    # a trivial 0 == 0 pass before any raft entries have been applied.
+    let survivorIdx = [0, 1, 2].filterIt(it != leaderIdx)
+    var caughtUp = false
+    let cuStart = getTime()
+    while getTime() - cuStart < initDuration(seconds = 10):
+      let c0 = countRows(nodes[survivorIdx[0]].clientPort)
+      let c1 = countRows(nodes[survivorIdx[1]].clientPort)
+      if c0 >= ackedAtKill and c0 == c1:
+        caughtUp = true
+        break
+      sleep(100)
+    if not caughtUp:
+      echo "survivors never reached equal row counts within 10s (",
+           countRows(nodes[survivorIdx[0]].clientPort), " vs ",
+           countRows(nodes[survivorIdx[1]].clientPort), ")"
+      dumpAll(nodes)
+      fail()
+      return
+    echo "survivors caught up: equal row counts"
+
+    # Assert B (durability): every acknowledged id must be present on BOTH
+    # survivors.
+    var acked: HashSet[int]
+    withLock wLock:
+      acked = toHashSet(wAcked)
+    for i in survivorIdx:
+      var ids: HashSet[int]
+      try:
+        ids = allIds(nodes[i].clientPort)
+      except CatchableError as e:
+        echo "ASSERT B FAILED: could not read ids from ", nodes[i].id,
+             ": ", e.msg
+        dumpAll(nodes)
+        fail()
+        return
+      let missing = acked - ids
+      if missing.len > 0:
+        echo "ASSERT B FAILED: ", nodes[i].id, " is missing ",
+             missing.len, " acknowledged ids"
+        dumpAll(nodes)
+        fail()
+        return
+      echo "durability: ", nodes[i].id, " contains all ",
+           acked.len, " acknowledged ids"
+
+    check acked.len >= AckTarget
+  finally:
+    if writerStarted:
+      withLock wLock:
+        wStop = true
+      joinThreads(writer)
+    deinitLock(wLock)
+    for n in nodes.mitems:
+      n.killNode()
+      if n.p != nil: n.p.close()
+      removeDir(n.dataDir)
+
+suite "Raft failover under load E2E":
+  test "committed writes survive a leader kill under sustained write load":
+    if not fileExists(BinaryPath):
+      if getEnv("CI").len > 0:
+        echo "[FAIL] ", BinaryPath, " missing under CI — build step broken?"
+        fail()
+      else:
+        echo "[SKIP] ", BinaryPath, " missing — run `nimble test` (builds the server first)"
+        skip()
+    else:
+      runFailoverLoadScenario()

tests/raft_tls_e2e_test.nim

@@ -0,0 +1,445 @@
+## Raft TLS E2E — real 3-node cluster with TLS on the raft transport.
+## Starts three actual build/baradadb processes with per-node self-signed
+## certs (BARADB_RAFT_TLS_ENABLED + CERT/KEY_FILE), asserts election and
+## replicated writes work over the encrypted transport, then starts a 4th
+## plaintext node pointed at the same peers and asserts it never becomes
+## leader (its frames are undecryptable) while the TLS cluster keeps
+## operating among its 3 members.
+## Process-management conventions follow tests/raft_writes_e2e_test.nim
+## (copying is the repo's e2e convention — do not factor out).
+##
+## NOTE: only the raft port is TLS here. The SQL client port stays
+## plaintext (BARADB_TLS_ENABLED is the client wire port — a different
+## feature), so the baradb_sqlite adaptor connects as usual.
+import std/unittest
+import std/osproc
+import std/os
+import std/strtabs
+import std/strutils
+import std/times
+import std/net
+import std/posix
+
+import ../adaptors/nim/baradb_sqlite as sqlite
+import barabadb/protocol/ssl
+
+const
+  BinaryPath = "./build/baradadb"
+  LeaderMarker = "became leader"
+
+type
+  NodeProc = object
+    id: string
+    clientPort: int
+    raftPort: int
+    p: Process
+    dataDir: string
+    output: string
+    alive: bool
+
+proc drainOutput(n: var NodeProc) =
+  ## Reads whatever the child has written so far. The pipe was set O_NONBLOCK
+  ## at start, so this never blocks — a hung read is impossible here.
+  var tmp: array[8192, char]
+  while true:
+    let count = posix.read(n.p.outputHandle.cint, tmp[0].addr, tmp.len)
+    if count <= 0: break
+    for i in 0 ..< count: n.output.add tmp[i]
+
+proc drainAll(nodes: var seq[NodeProc]) =
+  for n in nodes.mitems:
+    if n.p != nil: n.drainOutput()
+
+proc dumpAll(nodes: var seq[NodeProc]) =
+  ## Debuggability: on failure, everything the nodes said.
+  nodes.drainAll()
+  for n in nodes:
+    echo "===== output of ", n.id, " (port ", n.clientPort, ") ====="
+    echo n.output
+
+proc portOpen(port: int): bool =
+  var s: Socket
+  try:
+    s = newSocket()
+    s.connect("127.0.0.1", Port(port), timeout = 250)
+    s.close()
+    result = true
+  except CatchableError:
+    if s != nil: s.close()
+    result = false
+
+proc tlsHandshake(port: int): bool =
+  ## True when a real TLS client handshake completes against `port`.
+  ## Wire-level discriminator: against a plaintext peer it fails — either
+  ## immediately (SSL error on the garbage reply) or, when the peer swallows
+  ## our ClientHello and waits for more bytes, via the 3s recv timeout.
+  var ctx: SslContext
+  var s: Socket
+  try:
+    ctx = newContext(verifyMode = CVerifyNone)
+    s = newSocket()
+    var tv = Timeval(tvSec: posix.Time(3), tvUsec: Suseconds(0))
+    discard setsockopt(s.getFd(), SOL_SOCKET, SO_RCVTIMEO,
+                       addr tv, SockLen(sizeof(tv)))
+    s.connect("127.0.0.1", Port(port), timeout = 3000)
+    ctx.wrapConnectedSocket(s, handshakeAsClient)
+    result = true
+  except CatchableError:
+    result = false
+  finally:
+    if s != nil: s.close()
+    if ctx != nil: ctx.destroyContext()
+
+proc killNode(n: var NodeProc) =
+  if n.p != nil and n.alive:
+    try:
+      n.p.terminate()
+      discard n.p.waitForExit()
+    except CatchableError:
+      discard
+    n.alive = false
+
+proc leaderTerms(output: string): seq[int] =
+  ## All terms this node logged leadership for ("became leader for term T").
+  var pos = 0
+  while true:
+    let idx = output.find(LeaderMarker, pos)
+    if idx < 0: break
+    let tIdx = output.find("term ", idx)
+    if tIdx < 0: break
+    let numStart = tIdx + 5
+    var numEnd = numStart
+    while numEnd < output.len and output[numEnd] in Digits: inc numEnd
+    if numEnd > numStart:
+      result.add(parseInt(output[numStart ..< numEnd]))
+    pos = numEnd
+
+proc maxLeader(nodes: seq[NodeProc]): tuple[idx, term: int] =
+  ## Node that logged leadership for the highest term seen so far.
+  result = (-1, 0)
+  for i in 0 ..< nodes.len:
+    for t in leaderTerms(nodes[i].output):
+      if t > result.term: result = (i, t)
+
+proc drainFor(nodes: var seq[NodeProc], ms: int) =
+  let start = getTime()
+  while getTime() - start < initDuration(milliseconds = ms):
+    nodes.drainAll()
+    sleep(50)
+
+proc openClient(port: int): DbConn =
+  ## Connect with retries — the port may accept TCP before the DB is usable.
+  for i in 0 ..< 50:
+    try:
+      return open("127.0.0.1:" & $port, "", "", "default")
+    except CatchableError:
+      sleep(100)
+  raise newException(IOError, "cannot connect to port " & $port)
+
+proc waitForRow(port: int, table, name: string, deadlineSec: int): bool =
+  ## Poll SELECT on `port` until a row with `name` appears. Tolerates errors
+  ## (e.g. "unknown table" while schema has not been created yet) by retrying.
+  let db = openClient(port)
+  defer: db.close()
+  let start = getTime()
+  while getTime() - start < initDuration(seconds = deadlineSec):
+    try:
+      let rows = db.getAllRows(sql("SELECT * FROM " & table))
+      for row in rows:
+        if row.len >= 2 and row[1] == name:
+          return true
+    except CatchableError:
+      discard
+    sleep(100)
+  return false
+
+proc startNode(id: string, clientPort, raftPort: int, peers, clientPeers: string,
+               tlsEnabled: bool): NodeProc =
+  ## Boots one build/baradadb process. When tlsEnabled, a self-signed cert
+  ## (CN = node id) is generated into the node's data dir first.
+  let tstamp = getTime().toUnix.int
+  let dataDir = getTempDir() / "baradb_raft_tls_e2e_" & $tstamp & "_" & id
+  createDir(dataDir)
+  var env = newStringTable()
+  for key, val in envPairs():
+    env[key] = val
+  env["BARADB_PORT"] = $clientPort
+  env["BARADB_RAFT_ENABLED"] = "true"
+  env["BARADB_RAFT_PORT"] = $raftPort
+  env["BARADB_RAFT_NODE_ID"] = id
+  env["BARADB_RAFT_PEERS"] = peers
+  env["BARADB_RAFT_CLIENT_PEERS"] = clientPeers
+  env["BARADB_DATA_DIR"] = dataDir
+  env["BARADB_LOG_LEVEL"] = "info"
+  if tlsEnabled:
+    let (certFile, keyFile) = generateSelfSignedCert(dataDir, id)
+    doAssert certFile.len > 0 and keyFile.len > 0,
+      "openssl cert generation failed for " & id
+    env["BARADB_RAFT_TLS_ENABLED"] = "true"
+    env["BARADB_RAFT_TLS_CERT_FILE"] = certFile
+    env["BARADB_RAFT_TLS_KEY_FILE"] = keyFile
+  let p = startProcess(BinaryPath, env = env,
+                       options = {poStdErrToStdOut, poDaemon})
+  discard fcntl(p.outputHandle.cint, F_SETFL,
+                fcntl(p.outputHandle.cint, F_GETFL) or O_NONBLOCK)
+  NodeProc(id: id, clientPort: clientPort, raftPort: raftPort, p: p,
+           dataDir: dataDir, alive: true)
+
+proc runTlsScenario() =
+  ## Fatal phase failures dump all captured node output, record a test
+  ## failure, and return; cleanup happens in the finally below either way.
+  let tstamp = getTime().toUnix.int
+  # Port base per the T6 brief: distinct from raft_writes_e2e_test
+  # (46000+mod4000). Client ports are spaced by 10 because the server
+  # derives HTTP (port+440), WS (port+441) and gossip (raftPort+100)
+  # ports — consecutive client ports collide.
+  let cbase = 54000 + (tstamp mod 4000)
+  let rbase = cbase + 100
+  let peers = "n1@127.0.0.1:" & $(rbase + 1) &
+              ",n2@127.0.0.1:" & $(rbase + 2) &
+              ",n3@127.0.0.1:" & $(rbase + 3)
+  # SQL client ports for transparent leader write forwarding.
+  let clientPeers = "n1@127.0.0.1:" & $(cbase + 10) &
+                    ",n2@127.0.0.1:" & $(cbase + 20) &
+                    ",n3@127.0.0.1:" & $(cbase + 30)
+
+  var nodes: seq[NodeProc]
+  # Negative-case node: same peers, raft TLS DISABLED, own dir and ports
+  # (4th port in each base range). Started later, after the TLS cluster is
+  # up, so the positive assertions are not polluted by its noise.
+  var rogue: NodeProc
+
+  try:
+    # Node starts live inside the try so a raise from start #2/#3 still
+    # reaches the cleanup in the finally below.
+    for i in 1 .. 3:
+      nodes.add startNode("n" & $i, cbase + i * 10, rbase + i,
+                          peers, clientPeers, tlsEnabled = true)
+
+    # Readiness: all three client ports accept TCP connections (10s each).
+    for i in 0 ..< nodes.len:
+      let readyStart = getTime()
+      var ok = false
+      while getTime() - readyStart < initDuration(seconds = 10):
+        if portOpen(nodes[i].clientPort):
+          ok = true
+          break
+        sleep(100)
+      if not ok:
+        echo "node ", nodes[i].id, " never became ready"
+        dumpAll(nodes)
+        fail()
+        return
+
+    # Election over TLS: timeouts are 150-300ms, heartbeat 50ms — a leader
+    # should emerge within ~2s; 10s deadline for margin.
+    var elected = false
+    let electStart = getTime()
+    while getTime() - electStart < initDuration(seconds = 10):
+      nodes.drainAll()
+      if maxLeader(nodes).idx >= 0:
+        elected = true
+        break
+      sleep(50)
+    if not elected:
+      echo "no leader elected within 10s over TLS"
+      dumpAll(nodes)
+      fail()
+      return
+
+    # Settle and require stability, same as raft_writes_e2e_test.
+    nodes.drainFor(2000)
+    let (leaderIdx, leaderTerm) = maxLeader(nodes)
+    nodes.drainFor(1000)
+    let (stableIdx, stableTerm) = maxLeader(nodes)
+    if stableIdx != leaderIdx or stableTerm != leaderTerm:
+      echo "TLS cluster unstable: leadership moved from ", nodes[leaderIdx].id,
+           " (term ", leaderTerm, ") to ", nodes[stableIdx].id,
+           " (term ", stableTerm, ")"
+      dumpAll(nodes)
+      fail()
+      return
+    echo "TLS leader elected: ", nodes[leaderIdx].id, " (term ", leaderTerm, ")"
+
+    # Wire-level TLS assertion: a real client handshake must complete
+    # against every cluster node's raft port. Without this, the negative
+    # "plaintext node never becomes leader" check alone cannot distinguish
+    # TLS rejection from "can't win an election anyway" — if TLS wrapping
+    # were silently dropped from the transport, these handshakes raise and
+    # the suite fails.
+    for n in nodes.items:
+      if not tlsHandshake(n.raftPort):
+        echo "TLS handshake failed against raft port of ", n.id,
+             " (port ", n.raftPort, ") — raft transport not encrypted?"
+        dumpAll(nodes)
+        fail()
+        return
+    echo "TLS handshake verified against all 3 raft ports"
+
+    let followerIdx = (if leaderIdx == 0: 1 else: 0)
+
+    # Schema: CREATE TABLE goes through the raft "ddl" log (C3c) — this
+    # proves raft DDL replication works over the TLS transport.
+    block:
+      let db = openClient(nodes[leaderIdx].clientPort)
+      try:
+        db.exec(sql"CREATE TABLE tls_test (id INT PRIMARY KEY, name STRING)")
+      except CatchableError as e:
+        echo "leader CREATE TABLE failed: ", e.msg
+        dumpAll(nodes)
+        fail()
+        return
+      db.close()
+    echo "leader schema committed via raft ddl over TLS"
+
+    # Wait until the follower has applied CREATE TABLE (SELECT no longer
+    # errors with unknown table). Deadline 5s.
+    block:
+      let db = openClient(nodes[followerIdx].clientPort)
+      defer: db.close()
+      let start = getTime()
+      var ready = false
+      while getTime() - start < initDuration(seconds = 5):
+        try:
+          discard db.getAllRows(sql"SELECT * FROM tls_test")
+          ready = true
+          break
+        except CatchableError:
+          sleep(100)
+      if not ready:
+        echo "follower ", nodes[followerIdx].id,
+             " never applied CREATE TABLE within 5s"
+        dumpAll(nodes)
+        fail()
+        return
+    echo "schema replicated to follower ", nodes[followerIdx].id, " over TLS"
+
+    # Leader write: INSERT goes through the raft log and waits for majority
+    # commit before responding — expect success over TLS.
+    block:
+      let db = openClient(nodes[leaderIdx].clientPort)
+      try:
+        db.exec(sql"INSERT INTO tls_test (id, name) VALUES (1, 'tls-row')")
+      except CatchableError as e:
+        echo "leader INSERT failed: ", e.msg
+        dumpAll(nodes)
+        fail()
+        return
+      db.close()
+    echo "leader INSERT committed over TLS"
+
+    # Follower visibility: poll until the row shows up (5s deadline).
+    if not waitForRow(nodes[followerIdx].clientPort, "tls_test", "tls-row", 5):
+      echo "follower ", nodes[followerIdx].id,
+           " never saw the replicated row within 5s"
+      dumpAll(nodes)
+      fail()
+      return
+    echo "row replicated to follower ", nodes[followerIdx].id, " over TLS"
+
+    # Negative scenario: start the plaintext node pointed at the same peers.
+    # Its TLS-less frames are undecryptable to the cluster, so it can never
+    # collect votes and must never become leader.
+    rogue = startNode("n4", cbase + 40, rbase + 4, peers, clientPeers,
+                      tlsEnabled = false)
+    block:
+      let readyStart = getTime()
+      var ok = false
+      while getTime() - readyStart < initDuration(seconds = 10):
+        if portOpen(rogue.clientPort):
+          ok = true
+          break
+        sleep(100)
+      if not ok:
+        echo "plaintext node never became ready"
+        rogue.drainOutput()
+        echo "===== output of ", rogue.id, " (port ", rogue.clientPort, ") ====="
+        echo rogue.output
+        dumpAll(nodes)
+        fail()
+        return
+    echo "plaintext node n4 started against the TLS peers"
+
+    # Complementary wire-level negative: the rogue node's raft port speaks
+    # plaintext, so a TLS handshake against it must FAIL.
+    if tlsHandshake(rogue.raftPort):
+      echo "TLS handshake unexpectedly succeeded against plaintext n4 ",
+           "raft port ", rogue.raftPort
+      dumpAll(nodes)
+      fail()
+      return
+    echo "TLS handshake correctly fails against plaintext n4 raft port"
+
+    # Give n4 many election cycles (timeouts 150-300ms) to try its luck.
+    # The TLS cluster must keep operating among its 3 members meanwhile.
+    # Drain n4's pipe alongside the others — its connection-refused spam
+    # must not fill the pipe and block the child.
+    block:
+      let negStart = getTime()
+      while getTime() - negStart < initDuration(seconds = 6):
+        nodes.drainAll()
+        rogue.drainOutput()
+        sleep(50)
+
+    # Cluster still elects/operates: leader among the 3 TLS nodes accepts a
+    # write and it replicates to a follower, with n4 running.
+    nodes.drainAll()
+    let (leaderIdx2, leaderTerm2) = maxLeader(nodes)
+    if leaderIdx2 < 0:
+      echo "TLS cluster lost its leader after plaintext node joined"
+      dumpAll(nodes)
+      fail()
+      return
+    let followerIdx2 = (if leaderIdx2 == 0: 1 else: 0)
+    block:
+      let db = openClient(nodes[leaderIdx2].clientPort)
+      try:
+        db.exec(sql"INSERT INTO tls_test (id, name) VALUES (2, 'still-tls')")
+      except CatchableError as e:
+        echo "post-plaintext INSERT failed on ", nodes[leaderIdx2].id,
+             " (term ", leaderTerm2, "): ", e.msg
+        rogue.drainOutput()
+        echo "===== output of ", rogue.id, " (port ", rogue.clientPort, ") ====="
+        echo rogue.output
+        dumpAll(nodes)
+        fail()
+        return
+      db.close()
+    if not waitForRow(nodes[followerIdx2].clientPort, "tls_test", "still-tls", 5):
+      echo "post-plaintext row never replicated to ", nodes[followerIdx2].id
+      dumpAll(nodes)
+      fail()
+      return
+    echo "TLS cluster kept operating with plaintext node present"
+
+    # Final negative assertion: n4 never won an election.
+    rogue.drainOutput()
+    if rogue.output.contains(LeaderMarker):
+      echo "plaintext node became leader — TLS isolation broken"
+      echo "===== output of ", rogue.id, " (port ", rogue.clientPort, ") ====="
+      echo rogue.output
+      dumpAll(nodes)
+      fail()
+      return
+    echo "plaintext node never became leader (TLS rejection confirmed)"
+  finally:
+    for n in nodes.mitems:
+      n.killNode()
+      if n.p != nil: n.p.close()
+      removeDir(n.dataDir)
+    if rogue.p != nil:
+      rogue.killNode()
+      rogue.p.close()
+      removeDir(rogue.dataDir)
+
+suite "Raft TLS E2E":
+  test "3-node TLS cluster elects and replicates; plaintext node rejected":
+    if not fileExists(BinaryPath):
+      if getEnv("CI").len > 0:
+        echo "[FAIL] ", BinaryPath, " missing under CI — build step broken?"
+        fail()
+      else:
+        echo "[SKIP] ", BinaryPath, " missing — run `nimble test` (builds the server first)"
+        skip()
+    else:
+      runTlsScenario()

tests/raft_writes_e2e_test.nim

@@ -412,7 +412,11 @@ proc runWritesScenario() =
 suite "Raft replicated writes E2E":
   test "writes replicate, followers reject, failover resumes writes":
     if not fileExists(BinaryPath):
-      echo "[SKIP] ", BinaryPath, " missing — run `nimble test` (builds the server first)"
-      skip()
+      if getEnv("CI").len > 0:
+        echo "[FAIL] ", BinaryPath, " missing under CI — build step broken?"
+        fail()
+      else:
+        echo "[SKIP] ", BinaryPath, " missing — run `nimble test` (builds the server first)"
+        skip()
     else:
       runWritesScenario()

tests/test_all.nim

@@ -1,6 +1,7 @@
 ## BaraDB — Test Suite
 import std/unittest
 import std/tables
+import std/sets
 import std/strutils
 import std/os
 import std/asyncdispatch
@@ -8,11 +9,13 @@ import std/asyncnet
 import std/monotimes
 import std/base64
 import std/json
+import std/streams
 
 import barabadb/core/types
 import barabadb/core/mvcc
 import barabadb/core/deadlock
 import barabadb/core/config
+import barabadb/core/backup
 import barabadb/core/server
 import barabadb/core/columnar
 import barabadb/core/raft
@@ -1656,14 +1659,28 @@ suite "Replication":
     rm.connectReplica("r2")
     rm.connectReplica("r3")
 
+    # Unreachable replicas cannot ack — semi-sync must fail closed (return 0)
     let lsn = rm.writeLsn(@[1'u8])
-    check not rm.isFullyAcked(lsn)  # needs 2 acks
+    check lsn == 0
 
-    rm.ackLsn("r1", lsn)
-    check not rm.isFullyAcked(lsn)  # still needs 1 more
+    # No connected replicas → nothing to wait for; write succeeds
+    var rm2 = newReplicationManager(rmSemiSync, syncCount = 2)
+    rm2.addReplica(newReplica("r1", "10.0.0.1", 9472))
+    # not connected
+    let lsn2 = rm2.writeLsn(@[1'u8])
+    check lsn2 > 0
+    check rm2.isFullyAcked(lsn2)
 
-    rm.ackLsn("r2", lsn)
-    check rm.isFullyAcked(lsn)  # 2 acks received
+    # ackLsn bookkeeping still clears pendingAcks at the required quorum
+    var rm3 = newReplicationManager(rmSemiSync, syncCount = 2)
+    rm3.pendingAcks[1'u64] = initHashSet[string]()
+    rm3.pendingAcks[1'u64].incl("r1")
+    rm3.pendingAcks[1'u64].incl("r2")
+    check not rm3.isFullyAcked(1)
+    rm3.ackLsn("r1", 1)
+    check not rm3.isFullyAcked(1)
+    rm3.ackLsn("r2", 1)
+    check rm3.isFullyAcked(1)
 
   test "Replica status":
     var rm = newReplicationManager(rmAsync)
@@ -2607,6 +2624,613 @@ suite "Raft Network Transport":
       check replyMsg.kind == rmkRequestVoteReply
       check replyMsg.success
 
+suite "Raft InstallSnapshot Protocol":
+  test "InstallSnapshot fields survive serialize/deserialize round-trip":
+    let msg = RaftMessage(
+      kind: rmkInstallSnapshot,
+      term: 9,
+      senderId: "leader-1",
+      prevLogIndex: 42,   # snapshot base index
+      prevLogTerm: 7,     # snapshot base term
+      snapId: 3,
+      snapOffset: 4096,
+      snapData: @[byte 1, 2, 3, 250, 0, 17],
+      snapDone: true)
+    let decoded = deserializeRaftMessage(serialize(msg))
+    check decoded.kind == rmkInstallSnapshot
+    check decoded.term == 9
+    check decoded.senderId == "leader-1"
+    check decoded.prevLogIndex == 42
+    check decoded.prevLogTerm == 7
+    check decoded.snapId == 3
+    check decoded.snapOffset == 4096
+    check decoded.snapData == @[byte 1, 2, 3, 250, 0, 17]
+    check decoded.snapDone
+
+  test "InstallSnapshotReply fields survive serialize/deserialize round-trip":
+    let msg = RaftMessage(
+      kind: rmkInstallSnapshotReply,
+      term: 9,
+      senderId: "follower-2",
+      success: true,
+      matchIdx: 42,
+      snapId: 3,
+      snapOffset: 8192,
+      snapData: @[],
+      snapDone: false)
+    let decoded = deserializeRaftMessage(serialize(msg))
+    check decoded.kind == rmkInstallSnapshotReply
+    check decoded.term == 9
+    check decoded.senderId == "follower-2"
+    check decoded.success
+    check decoded.matchIdx == 42
+    check decoded.snapId == 3
+    check decoded.snapOffset == 8192
+    check decoded.snapData.len == 0
+    check not decoded.snapDone
+
+  test "old wire layout (no snapshot fields) deserializes with zero defaults":
+    # Manually serialize a message in the pre-InstallSnapshot layout:
+    # magic, version, kind, term, senderId, lastLogIndex, lastLogTerm,
+    # prevLogIndex, prevLogTerm, entries, leaderCommit, success, matchIdx.
+    let s = newStringStream()
+    s.write("RAFT")
+    s.write(1'u32)  # RaftProtoVersion
+    s.write(uint32(ord(rmkAppendEntries)))
+    s.write(5'u64)  # term
+    let sender = "old-leader"
+    s.write(uint32(sender.len))
+    s.writeData(sender[0].unsafeAddr, sender.len)
+    s.write(11'u64)  # lastLogIndex
+    s.write(4'u64)   # lastLogTerm
+    s.write(10'u64)  # prevLogIndex
+    s.write(4'u64)   # prevLogTerm
+    s.write(0'u32)   # entries count
+    s.write(10'u64)  # leaderCommit
+    s.write(char(1)) # success
+    s.write(10'u64)  # matchIdx
+    let strData = s.data
+    var buf = newSeq[byte](strData.len)
+    for i in 0 ..< strData.len:
+      buf[i] = byte(strData[i])
+    s.close()
+
+    let decoded = deserializeRaftMessage(buf)
+    check decoded.kind == rmkAppendEntries
+    check decoded.term == 5
+    check decoded.senderId == "old-leader"
+    check decoded.matchIdx == 10
+    check decoded.snapId == 0
+    check decoded.snapOffset == 0
+    check decoded.snapData.len == 0
+    check not decoded.snapDone
+
+  test "old message kinds still round-trip unchanged":
+    let msg = RaftMessage(kind: rmkRequestVote, term: 2, senderId: "cand",
+                          lastLogIndex: 5, lastLogTerm: 1)
+    let decoded = deserializeRaftMessage(serialize(msg))
+    check decoded.kind == rmkRequestVote
+    check decoded.term == 2
+    check decoded.senderId == "cand"
+    check decoded.lastLogIndex == 5
+    check decoded.lastLogTerm == 1
+    check decoded.snapId == 0
+    check decoded.snapOffset == 0
+    check decoded.snapData.len == 0
+    check not decoded.snapDone
+
+suite "Raft InstallSnapshot Receive":
+  test "follower assembles chunks, restores snapshot, resets state":
+    proc scenario() =
+      let tmp = getTempDir() / "baradb_snaprx_ok_" & $getCurrentProcessId()
+      removeDir(tmp)
+      createDir(tmp)
+      defer: removeDir(tmp)
+
+      # Real tar.gz fixture with a marker file
+      let srcDb = tmp / "srcdb"
+      createDir(srcDb)
+      writeFile(srcDb / "marker.txt", "snapshot-payload")
+      let archivePath = tmp / "snap.tar.gz"
+      check backupDataDir(srcDb, archivePath)
+      let archiveBytes = readFile(archivePath)
+      check archiveBytes.len > 0
+
+      let raftDir = tmp / "raft"
+      var node = newRaftNode("follower-1", @[], dataDir = raftDir)
+      node.currentTerm = 5
+      node.log.add(LogEntry(term: 3, index: 10, command: "put", data: @[byte 1]))
+      node.commitIndex = 10
+
+      var gotPath = ""
+      var gotBaseIndex = 0'u64
+      var gotBaseTerm = 0'u64
+      node.restoreSnapshot = proc(p: string, bi: uint64, bt: uint64): bool {.gcsafe.} =
+        gotPath = p
+        gotBaseIndex = bi
+        gotBaseTerm = bt
+        # Assembled archive must match the original byte-for-byte
+        result = readFile(p) == archiveBytes
+
+      let half = archiveBytes.len div 2
+      let reply1 = node.handleInstallSnapshot(RaftMessage(
+        kind: rmkInstallSnapshot, term: 5, senderId: "leader-1",
+        prevLogIndex: 40, prevLogTerm: 4,
+        snapId: 7, snapOffset: 0,
+        snapData: cast[seq[byte]](archiveBytes[0 ..< half]), snapDone: false))
+      check reply1.kind == rmkInstallSnapshotReply
+      check reply1.success
+
+      let reply2 = node.handleInstallSnapshot(RaftMessage(
+        kind: rmkInstallSnapshot, term: 5, senderId: "leader-1",
+        prevLogIndex: 40, prevLogTerm: 4,
+        snapId: 7, snapOffset: uint64(half),
+        snapData: cast[seq[byte]](archiveBytes[half .. ^1]), snapDone: true))
+      check reply2.success
+      check reply2.matchIdx == 40
+
+      check gotPath.len > 0
+      check "snap_incoming" in gotPath
+      check gotBaseIndex == 40
+      check gotBaseTerm == 4
+      check node.lastSnapshotIndex == 40
+      check node.lastSnapshotTerm == 4
+      check node.commitIndex == 40
+      check node.lastApplied == 40
+      check node.log.len == 0
+
+      # State was persisted: a fresh node on the same dir sees the snapshot base
+      let reloaded = newRaftNode("follower-1", @[], dataDir = raftDir)
+      check reloaded.lastSnapshotIndex == 40
+      check reloaded.lastSnapshotTerm == 4
+      check reloaded.log.len == 0
+    scenario()
+
+  test "failed restore leaves state untouched and removes temp file":
+    proc scenario() =
+      let tmp = getTempDir() / "baradb_snaprx_fail_" & $getCurrentProcessId()
+      removeDir(tmp)
+      createDir(tmp)
+      defer: removeDir(tmp)
+
+      let raftDir = tmp / "raft"
+      var node = newRaftNode("follower-1", @[], dataDir = raftDir)
+      node.currentTerm = 5
+      node.log.add(LogEntry(term: 2, index: 3, command: "put", data: @[byte 9]))
+      node.restoreSnapshot = proc(p: string, bi: uint64, bt: uint64): bool {.gcsafe.} =
+        false
+
+      let reply = node.handleInstallSnapshot(RaftMessage(
+        kind: rmkInstallSnapshot, term: 5, senderId: "leader-1",
+        prevLogIndex: 8, prevLogTerm: 2,
+        snapId: 1, snapOffset: 0,
+        snapData: @[byte 1, 2, 3], snapDone: true))
+      check not reply.success
+      check node.lastSnapshotIndex == 0
+      check node.lastSnapshotTerm == 0
+      check node.commitIndex == 0
+      check node.log.len == 1
+      check not fileExists(raftDir / "snap_incoming" / "snap_1.tar.gz")
+    scenario()
+
+  test "oversized chunk is rejected":
+    let tmp = getTempDir() / "baradb_snaprx_cap_" & $getCurrentProcessId()
+    removeDir(tmp)
+    createDir(tmp)
+    defer: removeDir(tmp)
+
+    var node = newRaftNode("follower-1", @[], dataDir = tmp / "raft")
+    node.currentTerm = 1
+    node.snapChunkBytes = 4
+    let reply = node.handleInstallSnapshot(RaftMessage(
+      kind: rmkInstallSnapshot, term: 1, senderId: "leader-1",
+      prevLogIndex: 1, prevLogTerm: 1,
+      snapId: 1, snapOffset: 0,
+      snapData: @[byte 1, 2, 3, 4, 5], snapDone: false))
+    check not reply.success
+    check node.snapIncomingId == 0
+
+  test "out-of-order offset is rejected and assembly restarts on new snapId":
+    let tmp = getTempDir() / "baradb_snaprx_off_" & $getCurrentProcessId()
+    removeDir(tmp)
+    createDir(tmp)
+    defer: removeDir(tmp)
+
+    let raftDir = tmp / "raft"
+    var node = newRaftNode("follower-1", @[], dataDir = raftDir)
+    node.currentTerm = 1
+    let ok1 = node.handleInstallSnapshot(RaftMessage(
+      kind: rmkInstallSnapshot, term: 1, senderId: "leader-1",
+      prevLogIndex: 2, prevLogTerm: 1,
+      snapId: 3, snapOffset: 0,
+      snapData: @[byte 65, 66], snapDone: false))
+    check ok1.success
+    # Gap: offset 5 while only 2 bytes assembled
+    let bad = node.handleInstallSnapshot(RaftMessage(
+      kind: rmkInstallSnapshot, term: 1, senderId: "leader-1",
+      prevLogIndex: 2, prevLogTerm: 1,
+      snapId: 3, snapOffset: 5,
+      snapData: @[byte 67], snapDone: false))
+    check not bad.success
+    check node.snapIncomingId == 0
+
+suite "Raft InstallSnapshot Send":
+  test "two consecutive floor rejects queue a snapshot send":
+    var node = newRaftNode("leader", @["peer-1"])
+    node.currentTerm = 5
+    node.state = rsLeader
+    node.lastSnapshotIndex = 100
+    node.lastSnapshotTerm = 4
+    node.nextIndex["peer-1"] = 101
+    node.matchIndex["peer-1"] = 0
+
+    let reject = RaftMessage(kind: rmkAppendEntriesReply, term: 5,
+                             senderId: "peer-1", success: false)
+    node.handleAppendReply("peer-1", reject)
+    check node.snapRejectStreak["peer-1"] == 1
+    check "peer-1" notin node.snapPending
+    check node.nextIndex["peer-1"] == 101  # pinned at the compaction floor
+
+    node.handleAppendReply("peer-1", reject)
+    check node.snapRejectStreak["peer-1"] == 2
+    check "peer-1" in node.snapPending
+    check node.nextIndex["peer-1"] == 101
+
+  test "non-floor reject decrements nextIndex without touching the streak":
+    var node = newRaftNode("leader", @["peer-1"])
+    node.currentTerm = 5
+    node.state = rsLeader
+    node.lastSnapshotIndex = 100
+    node.lastSnapshotTerm = 4
+    node.nextIndex["peer-1"] = 105
+
+    node.handleAppendReply("peer-1", RaftMessage(
+      kind: rmkAppendEntriesReply, term: 5, senderId: "peer-1", success: false))
+    check node.nextIndex["peer-1"] == 104
+    check "peer-1" notin node.snapRejectStreak
+    check "peer-1" notin node.snapPending
+
+  test "successful AppendEntries reply resets the streak and cancels a pending snapshot":
+    var node = newRaftNode("leader", @["peer-1"])
+    node.currentTerm = 5
+    node.state = rsLeader
+    node.lastSnapshotIndex = 100
+    node.lastSnapshotTerm = 4
+    node.nextIndex["peer-1"] = 101
+    node.matchIndex["peer-1"] = 0
+    node.snapRejectStreak["peer-1"] = 1
+    node.snapPending.incl("peer-1")
+
+    node.handleAppendReply("peer-1", RaftMessage(
+      kind: rmkAppendEntriesReply, term: 5, senderId: "peer-1",
+      success: true, matchIdx: 101))
+    check "peer-1" notin node.snapRejectStreak
+    check "peer-1" notin node.snapPending
+    check node.matchIndex["peer-1"] == 101
+    check node.nextIndex["peer-1"] == 102
+
+  test "commit requires strict majority for even-sized clusters":
+    ## Regression: the commit quorum used (N+1) div 2, which for a 4-node
+    ## cluster commits at 2/4 (a minority). Strict majority is N div 2 + 1.
+    var node = newRaftNode("leader", @["p1", "p2", "p3"])
+    node.currentTerm = 5
+    node.state = rsLeader
+    let e = node.appendLog("put", cast[seq[byte]]("k\x00v"))
+    check e.index == 1
+    check e.term == 5
+    node.nextIndex["p1"] = 2
+    node.nextIndex["p2"] = 2
+    node.nextIndex["p3"] = 2
+    # Leader + 1 peer (count=2) is NOT a majority of 4.
+    node.handleAppendReply("p1", RaftMessage(
+      kind: rmkAppendEntriesReply, term: 5, senderId: "p1",
+      success: true, matchIdx: 1))
+    check node.commitIndex == 0
+    # Leader + 2 peers (count=3) IS a strict majority of 4 -> commits.
+    node.handleAppendReply("p2", RaftMessage(
+      kind: rmkAppendEntriesReply, term: 5, senderId: "p2",
+      success: true, matchIdx: 1))
+    check node.commitIndex == 1
+
+  test "InstallSnapshotReply success advances match/next index and clears streak":
+    var node = newRaftNode("leader", @["peer-1"])
+    node.currentTerm = 5
+    node.state = rsLeader
+    node.lastSnapshotIndex = 100
+    node.lastSnapshotTerm = 4
+    node.nextIndex["peer-1"] = 101
+    node.matchIndex["peer-1"] = 0
+    node.snapRejectStreak["peer-1"] = 2
+    node.snapPending.incl("peer-1")
+
+    node.handleInstallSnapshotReply("peer-1", RaftMessage(
+      kind: rmkInstallSnapshotReply, term: 5, senderId: "peer-1",
+      success: true, matchIdx: 100))
+    check node.matchIndex["peer-1"] == 100
+    check node.nextIndex["peer-1"] == 101
+    check "peer-1" notin node.snapRejectStreak
+    check "peer-1" notin node.snapPending
+
+  test "InstallSnapshotReply failure leaves leader state untouched":
+    var node = newRaftNode("leader", @["peer-1"])
+    node.currentTerm = 5
+    node.state = rsLeader
+    node.lastSnapshotIndex = 100
+    node.lastSnapshotTerm = 4
+    node.nextIndex["peer-1"] = 101
+    node.matchIndex["peer-1"] = 0
+    node.snapRejectStreak["peer-1"] = 2
+
+    node.handleInstallSnapshotReply("peer-1", RaftMessage(
+      kind: rmkInstallSnapshotReply, term: 5, senderId: "peer-1",
+      success: false, matchIdx: 0))
+    check node.matchIndex["peer-1"] == 0
+    check node.nextIndex["peer-1"] == 101
+    check node.snapRejectStreak["peer-1"] == 2
+
+  test "intermediate chunk replies are ignored; final reply advances state":
+    var node = newRaftNode("leader", @["peer-1"])
+    node.currentTerm = 5
+    node.state = rsLeader
+    node.lastSnapshotIndex = 100
+    node.lastSnapshotTerm = 4
+    node.nextIndex["peer-1"] = 101
+    node.matchIndex["peer-1"] = 0
+    node.snapRejectStreak["peer-1"] = 2
+    node.snapPending.incl("peer-1")
+
+    # Intermediate chunk ack: follower replies success=true with its OLD
+    # lastSnapshotIndex (40 < our 100). Leader state must not move.
+    node.handleInstallSnapshotReply("peer-1", RaftMessage(
+      kind: rmkInstallSnapshotReply, term: 5, senderId: "peer-1",
+      success: true, matchIdx: 40))
+    check node.matchIndex["peer-1"] == 0
+    check node.nextIndex["peer-1"] == 101
+    check node.snapRejectStreak["peer-1"] == 2
+    check "peer-1" in node.snapPending
+
+    # Final reply: follower adopted the snapshot base (matchIdx == 100).
+    node.handleInstallSnapshotReply("peer-1", RaftMessage(
+      kind: rmkInstallSnapshotReply, term: 5, senderId: "peer-1",
+      success: true, matchIdx: 100))
+    check node.matchIndex["peer-1"] == 100
+    check node.nextIndex["peer-1"] == 101
+    check "peer-1" notin node.snapRejectStreak
+    check "peer-1" notin node.snapPending
+
+  test "InstallSnapshotReply term handling matches AppendEntriesReply":
+    var node = newRaftNode("leader", @["peer-1"])
+    node.currentTerm = 5
+    node.state = rsLeader
+    node.lastSnapshotIndex = 100
+    node.nextIndex["peer-1"] = 101
+    node.matchIndex["peer-1"] = 0
+
+    # Stale term: ignored entirely
+    node.handleInstallSnapshotReply("peer-1", RaftMessage(
+      kind: rmkInstallSnapshotReply, term: 4, senderId: "peer-1",
+      success: true, matchIdx: 100))
+    check node.matchIndex["peer-1"] == 0
+    check node.state == rsLeader
+
+    # Higher term: step down
+    node.handleInstallSnapshotReply("peer-1", RaftMessage(
+      kind: rmkInstallSnapshotReply, term: 7, senderId: "peer-1",
+      success: true, matchIdx: 100))
+    check node.state == rsFollower
+    check node.currentTerm == 7
+
+  test "floor rejects trigger sendSnapshot end-to-end via processMessage":
+    proc scenario() =
+      let tmp = getTempDir() / "baradb_snaptx_e2e_" & $getCurrentProcessId()
+      removeDir(tmp)
+      createDir(tmp)
+      defer: removeDir(tmp)
+
+      var payload = ""
+      for i in 0 ..< 200:
+        payload.add(char(32 + (i mod 90)))
+
+      var leader = newRaftNode("leader", @["peer-1"], raftPort = 29331,
+                               dataDir = tmp / "raft-l")
+      createDir(tmp / "raft-l")  # newRaftNode only reads; sendSnapshot writes here
+      leader.currentTerm = 5
+      leader.state = rsLeader
+      leader.lastSnapshotIndex = 100
+      leader.lastSnapshotTerm = 4
+      leader.nextIndex["peer-1"] = 101
+      leader.matchIndex["peer-1"] = 0
+      leader.snapChunkBytes = 64  # 200 bytes -> 4 chunks
+      leader.peerAddrs["peer-1"] = ("127.0.0.1", 29332)
+      var buildCalls = 0
+      leader.buildSnapshot = proc(destPath: string): bool {.gcsafe.} =
+        inc buildCalls
+        check "snap_out_100" in destPath
+        writeFile(destPath, payload)
+        true
+
+      var follower = newRaftNode("peer-1", @["leader"], raftPort = 29332,
+                                 dataDir = tmp / "raft-f")
+      follower.currentTerm = 1
+      var gotBaseIndex = 0'u64
+      var gotBaseTerm = 0'u64
+      follower.restoreSnapshot = proc(p: string, bi: uint64,
+                                      bt: uint64): bool {.gcsafe.} =
+        gotBaseIndex = bi
+        gotBaseTerm = bt
+        # sendSnapshot now gzips the tar off the event loop, so the assembled
+        # archive is gzip-compressed; decompress before comparing the bytes.
+        let raw = p & ".raw"
+        defer:
+          if fileExists(raw): removeFile(raw)
+        if not gunzipFile(p, raw):
+          return false
+        result = readFile(raw) == payload
+
+      let netL = newRaftNetwork(leader)
+      let netF = newRaftNetwork(follower)
+      asyncCheck netF.run()
+      waitFor sleepAsync(50)
+
+      # Two floor-level rejects through the real message path; the second one
+      # must trigger an async snapshot send (leader itself never listens).
+      let reject = RaftMessage(kind: rmkAppendEntriesReply, term: 5,
+                               senderId: "peer-1", success: false)
+      waitFor netL.processMessage(reject)
+      check "peer-1" notin leader.snapPending
+      waitFor netL.processMessage(reject)
+
+      var waited = 0
+      while follower.lastSnapshotIndex != 100 and waited < 3000:
+        waitFor sleepAsync(50)
+        waited += 50
+
+      netF.stop()
+      waitFor sleepAsync(50)
+
+      check buildCalls == 1
+      check follower.lastSnapshotIndex == 100
+      check follower.lastSnapshotTerm == 4
+      check gotBaseIndex == 100
+      check gotBaseTerm == 4
+      # Temp archives (uncompressed tar + compressed .tar.gz) cleaned up after
+      # the transfer
+      check not fileExists(tmp / "raft-l" / "snap_out_100.tar.gz")
+      check not fileExists(tmp / "raft-l" / "snap_out_100.tar")
+    scenario()
+
+  test "sendSnapshot single-flight guard skips a concurrent send":
+    let tmp = getTempDir() / "baradb_snaptx_guard_" & $getCurrentProcessId()
+    removeDir(tmp)
+    createDir(tmp)
+    defer: removeDir(tmp)
+
+    var node = newRaftNode("leader", @["peer-1"], dataDir = tmp / "raft")
+    node.currentTerm = 5
+    node.state = rsLeader
+    node.lastSnapshotIndex = 100
+    node.lastSnapshotTerm = 4
+    var buildCalls = 0
+    node.buildSnapshot = proc(destPath: string): bool {.gcsafe.} =
+      inc buildCalls
+      writeFile(destPath, "x")
+      true
+
+    let net = newRaftNetwork(node)
+    node.snapSending.incl("peer-1")  # a send is already in flight
+    waitFor net.sendSnapshot("peer-1")
+    check buildCalls == 0
+
+  test "sendSnapshot skips when there is no compacted snapshot":
+    let tmp = getTempDir() / "baradb_snaptx_zero_" & $getCurrentProcessId()
+    removeDir(tmp)
+    createDir(tmp)
+    defer: removeDir(tmp)
+
+    var node = newRaftNode("leader", @["peer-1"], dataDir = tmp / "raft")
+    node.currentTerm = 5
+    node.state = rsLeader
+    # lastSnapshotIndex == 0: snapId 0 can never be received by a follower
+    var buildCalls = 0
+    node.buildSnapshot = proc(destPath: string): bool {.gcsafe.} =
+      inc buildCalls
+      true
+
+    let net = newRaftNetwork(node)
+    waitFor net.sendSnapshot("peer-1")
+    check buildCalls == 0
+    check "peer-1" notin node.snapSending
+
+suite "Raft TLS Transport":
+  test "2-node election over TLS":
+    let certDir = getTempDir() / "baradb_test_raft_tls"
+    let (certPath, keyPath) = generateSelfSignedCert(certDir, "raft-tls.local")
+    if certPath.len == 0:
+      skip()  # openssl unavailable
+    else:
+      let tls = newTLSContext(newTLSConfig(certPath, keyPath))
+      var n1 = newRaftNode("n1", @["n2"], raftPort = 29301)
+      var n2 = newRaftNode("n2", @["n1"], raftPort = 29302)
+      n1.electionTimeout = 150
+      n2.electionTimeout = 350
+      n1.peerAddrs["n2"] = ("127.0.0.1", 29302)
+      n2.peerAddrs["n1"] = ("127.0.0.1", 29301)
+
+      let net1 = newRaftNetwork(n1, tls)
+      let net2 = newRaftNetwork(n2, tls)
+
+      asyncCheck net1.run()
+      asyncCheck net2.run()
+      waitFor sleepAsync(50)
+
+      # No manual ticks — timerLoop drives the election over TLS.
+      var leaderCount = 0
+      var waited = 0
+      while waited < 3000:
+        leaderCount = 0
+        if n1.isLeader: inc leaderCount
+        if n2.isLeader: inc leaderCount
+        if leaderCount == 1: break
+        waitFor sleepAsync(100)
+        waited += 100
+
+      net1.stop()
+      net2.stop()
+      waitFor sleepAsync(50)
+
+      check leaderCount == 1
+
+  test "plaintext dial to a TLS raft port has no protocol effect":
+    let certDir = getTempDir() / "baradb_test_raft_tls"
+    let (certPath, keyPath) = generateSelfSignedCert(certDir, "raft-tls.local")
+    if certPath.len == 0:
+      skip()  # openssl unavailable
+    else:
+      let tls = newTLSContext(newTLSConfig(certPath, keyPath))
+      var n = newRaftNode("srv", @["cli"], raftPort = 29311)
+      n.electionTimeout = 60000  # keep the server passive during the test
+      n.peerAddrs["cli"] = ("127.0.0.1", 29312)
+      let net = newRaftNetwork(n, tls)
+      asyncCheck net.run()
+      waitFor sleepAsync(50)
+
+      let termBefore = n.currentTerm
+
+      # A plaintext client sends a perfectly valid serialized raft frame; the
+      # bytes fail the TLS handshake, so nothing reaches the state machine.
+      let voteReq = RaftMessage(kind: rmkRequestVote, term: 42, senderId: "cli")
+      let data = serialize(voteReq)
+      var frame = newSeq[byte](4 + data.len)
+      frame[0] = byte(data.len shr 24)
+      frame[1] = byte(data.len shr 16)
+      frame[2] = byte(data.len shr 8)
+      frame[3] = byte(data.len)
+      for i in 0 ..< data.len:
+        frame[4 + i] = data[i]
+
+      let client = newAsyncSocket()
+      waitFor client.connect("127.0.0.1", Port(29311))
+      try:
+        waitFor client.send(cast[string](frame))
+      except CatchableError:
+        discard
+      waitFor sleepAsync(300)
+
+      # The server must have dropped the connection after the failed handshake.
+      var connectionDropped = false
+      try:
+        connectionDropped = (waitFor client.recv(1)).len == 0
+      except CatchableError:
+        connectionDropped = true
+      client.close()
+      net.stop()
+      waitFor sleepAsync(50)
+
+      check n.state == rsFollower
+      check n.currentTerm == termBefore
+      check n.votedFor == ""
+      check connectionDropped
+
 suite "Raft SQL Write Path":
   test "leader append+commit wait round-trips through applyCommand":
     proc scenario() =
@@ -2670,7 +3294,7 @@ suite "Raft SQL Write Path":
 
         # Server-side leader write path: append + wait for majority commit
         let (ok, errMsg) = waitFor appendWriteToRaft(leader,
-          @[("users.1", cast[seq[byte]]("alice"))], timeoutMs = 3000)
+          @[("users.1", cast[seq[byte]]("alice"), false)], timeoutMs = 3000)
         check ok
         if not ok: echo "appendWriteToRaft failed: ", errMsg
 
@@ -2760,6 +3384,53 @@ suite "Raft SQL Write Path":
     check n.lastSnapshotIndex == 0
     check n.log.len == 15
 
+  test "leader compactLog unpins from a stale peer (never replied, stale window exceeded)":
+    var n = newRaftNode("n1", @["n2"], raftPort = 29123)
+    n.logMaxEntries = 5
+    n.raftPeerStaleMs = 1000
+    n.becomeLeader()
+    n.matchIndex["n2"] = 0  # peer never caught up
+    # Last successful reply is long past the stale window.
+    n.matchIndexSeenMs["n2"] = getMonoTime().ticks() div 1_000_000 - 60_000
+    for i in 1 .. 15:
+      discard n.appendLog("put", cast[seq[byte]]("x"))
+      n.commitIndex = uint64(i)
+      n.lastApplied = uint64(i)
+    n.compactLog()
+    # Stale peer excluded from minMatch — compaction runs through lastApplied.
+    check n.lastSnapshotIndex == 15
+    check n.log.len == 0
+
+  test "leader compactLog still pins at a recently-responsive peer matchIndex":
+    var n = newRaftNode("n1", @["n2"], raftPort = 29124)
+    n.logMaxEntries = 5
+    n.raftPeerStaleMs = 1000
+    n.becomeLeader()
+    n.matchIndex["n2"] = 3
+    n.matchIndexSeenMs["n2"] = getMonoTime().ticks() div 1_000_000  # just replied
+    for i in 1 .. 15:
+      discard n.appendLog("put", cast[seq[byte]]("x"))
+      n.commitIndex = uint64(i)
+      n.lastApplied = uint64(i)
+    n.compactLog()
+    check n.lastSnapshotIndex == 3
+    check n.log.len == 12
+
+  test "leader compactLog respects grace window for an unreplied peer":
+    var n = newRaftNode("n1", @["n2"], raftPort = 29125)
+    n.logMaxEntries = 5
+    n.raftPeerStaleMs = 30000
+    n.becomeLeader()  # matchIndexSeenMs initialized to now — within grace
+    n.matchIndex["n2"] = 0  # peer has not replied yet
+    for i in 1 .. 15:
+      discard n.appendLog("put", cast[seq[byte]]("x"))
+      n.commitIndex = uint64(i)
+      n.lastApplied = uint64(i)
+    n.compactLog()
+    # Grace window still active — peer pins compaction at matchIndex 0.
+    check n.lastSnapshotIndex == 0
+    check n.log.len == 15
+
   test "appendDdlToRaft fails when node is not leader":
     var n = newRaftNode("n1", @["n2"], raftPort = 29113)
     let (ok, err) = waitFor appendDdlToRaft(n,
@@ -2782,7 +3453,7 @@ suite "Raft SQL Write Path":
     var n = newRaftNode("n1", @["n2"], raftPort = 29111)
     # Still a follower — appendLog returns index 0.
     let (ok, err) = waitFor appendWriteToRaft(n,
-      @[("k", cast[seq[byte]]("v"))], timeoutMs = 200)
+      @[("k", cast[seq[byte]]("v"), false)], timeoutMs = 200)
     check not ok
     check "lost leadership" in err
 
@@ -2791,7 +3462,7 @@ suite "Raft SQL Write Path":
     var n = newRaftNode("n1", @["n2", "n3"], raftPort = 29112)
     n.becomeLeader()
     let (ok, err) = waitFor appendWriteToRaft(n,
-      @[("k", cast[seq[byte]]("v"))], timeoutMs = 300)
+      @[("k", cast[seq[byte]]("v"), false)], timeoutMs = 300)
     check not ok
     check "raft commit timeout" in err
 
@@ -2816,6 +3487,43 @@ suite "Raft SQL Write Path":
     let (found, _) = db.get("t.id=1")
     check not found
 
+  test "REP receiver chain (encode -> decode -> apply) maintains indexes":
+    ## Mirrors server.nim's legacy REP handler: decodeRepPayload decides the op,
+    ## then applyReplicatedPut/Delete keep secondary indexes consistent. Guards
+    ## the wiring the receiver relies on — an indexed put must populate the
+    ## B-tree and a PK-only put (empty value) must apply as a put, not vanish.
+    var testDir = getTempDir() / "baradb_rep_recv_idx_" & $getCurrentProcessId() & "_" & $getMonoTime().ticks
+    createDir(testDir)
+    defer: removeDir(testDir)
+    var db = newLSMTree(testDir)
+    var ctx = qexec.newExecutionContext(db)
+    discard qexec.executeQuery(ctx, parse(
+      "CREATE TABLE t (id INT PRIMARY KEY, name STRING)"))
+    discard qexec.executeQuery(ctx, parse(
+      "CREATE INDEX idx_name ON t (name)"))
+    # Leader ships an indexed put; the receiver decodes and applies it.
+    let put = decodeRepPayload(
+      encodeRepPayload(false, "t.id=1", cast[seq[byte]]("name=alice")))
+    check put.op == ropPut
+    if put.op == ropPut:
+      applyReplicatedPut(ctx, put.key, put.value)
+    check ctx.btrees["t.name"].get("alice").len >= 1
+    # A PK-only put (empty value) must apply as a put, not a delete.
+    let pk = decodeRepPayload(encodeRepPayload(false, "t.id=2", @[]))
+    check pk.op == ropPut
+    if pk.op == ropPut:
+      applyReplicatedPut(ctx, pk.key, pk.value)
+    let (foundPk, _) = db.get("t.id=2")
+    check foundPk
+    # Leader ships a delete; the receiver drops the row and the index entry.
+    let del = decodeRepPayload(encodeRepPayload(true, "t.id=1", @[]))
+    check del.op == ropDelete
+    if del.op == ropDelete:
+      applyReplicatedDelete(ctx, del.key)
+    check ctx.btrees["t.name"].get("alice").len == 0
+    let (foundDel, _) = db.get("t.id=1")
+    check not foundDel
+
   test "applyReplicatedPut updates in-memory graphs":
     var testDir = getTempDir() / "baradb_raft_apply_g_" & $getCurrentProcessId() & "_" & $getMonoTime().ticks
     createDir(testDir)
@@ -3590,6 +4298,21 @@ suite "Window Functions":
       if row["name"] == "Bob":
         check row["last_sal"] == "90000"
 
+  test "SUM/AVG/COUNT window aggregates over a running frame":
+    ## Default frame is UNBOUNDED PRECEDING .. CURRENT ROW (lower.nim).
+    let r = qexec.executeQuery(ctx, parse(
+      "SELECT name, salary, SUM(salary) OVER (ORDER BY salary) AS running, COUNT(*) OVER (ORDER BY salary) AS cnt FROM employees"))
+    check r.success
+    check r.rows.len == 5
+    for row in r.rows:
+      if row["name"] == "Charlie":
+        check parseFloat($row["running"]) == 70000.0
+        check $row["cnt"] == "1"
+      if row["name"] == "Eve":
+        # 70000+75000+80000+90000+95000
+        check parseFloat($row["running"]) == 410000.0
+        check $row["cnt"] == "5"
+
 suite "GROUP BY Aggregates":
   var db: LSMTree
   var ctx: qexec.ExecutionContext