fix: audit batch 2 — semi-sync, DISTINCT, set ops, MERGE, storage hardening

Address 12 deep-audit findings: semi-sync fail-closed on partial ack,
COUNT/SUM/AVG(DISTINCT), UNION/INTERSECT/EXCEPT dedup, MERGE THEN DELETE,
WAL torn-record recovery, MVCC/checkpoint/flush/compaction/WAL rewrite
safety, mmap overflow bounds; remove stray protocol/scram ELF.

.gitignore

@@ -71,5 +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

@@ -3,7 +3,7 @@
 > 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 (батч 1): 5 | Остават: 23**
+> **Общо: ~28 находки | Поправени: 17 (батч 1: 5 + батч 2: 12, вкл. hygiene) | Остават: 12**
 
 ---
 
@@ -17,54 +17,56 @@
 | 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 |
 
-**Верификация:** `baradadb` build чист; `tests/test_all.nim` (пълен suite) и `tests/bugfix_test.nim` минават без `[FAILED]`.
+## 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).
 
 ---
 
-## Остават (23)
+## Остават (12)
 
-### 🟠 HIGH (8)
+### 🟠 HIGH (2)
 
 | # | Проблем | Файл | Предложен fix |
 |---|---------|------|---------------|
 | H2 | **TLS client връзките между възли не верифицират сертификата** — `forwardQueryToLeader` ползва `verifyMode = CVerifyNone` → MITM на клъстър линка. Raft client dials са със същия default (`raftTlsVerifyPeer: false`). | `core/server.nim:70` | Verify peer cert срещу CA при client handshake (fail-closed при enabled TLS) |
-| H3 | **Semi-sync репликация връща durable LSN при partial/zero ack** — `rmSync` връща 0 при partial ack, но `rmSemiSync` връща LSN безусловно (само debug echo при 0 acks). | `core/replication.nim:217` | `if syncReplicaCount > 0 and ackCount < syncReplicaCount: return 0` |
+| H10 | **B-tree `remove` separator convention** — audit: `splitChild` left-max vs `removeRec` right-min. Naive left-max rewrite of separators/borrows **fails** `prop_test` interleaved insert/remove; needs careful multi-level fix + more targeted repro first. | `storage/btree.nim:377` | Repro + full-tree separator invariant; keep borrow/merge/search consistent |
-| H6 | **`COUNT/SUM/AVG(DISTINCT)` игнорира DISTINCT** — `funcDistinct` се set-ва в парсера (BUG-017), но `aggDistinct` никога не се копира/чете; няма dedup в aggregate пътищата. | `query/exec/lower.nim:116`, `plan_exec.nim` | Копирай `aggDistinct = node.funcDistinct`; dedup чрез `HashSet[string]` преди count/sum/avg |
-| H7 | **`UNION/INTERSECT/EXCEPT` (без ALL) чупят с KeyError** — dedup ключът чете `row["$value"]`, но projected редове нямат този ключ. Само `UNION ALL` работи. | `query/executor.nim:459` | Dedup ключ от projected колоните (join `valueToString` по ред на `cols`), не `row["$value"]` |
-| H8 | **`MERGE ... WHEN MATCHED THEN DELETE` / `AND <cond>` не се изпълняват** — AST/parser полетата (BUG-032) съществуват, но executor-ът не ги реферира; DELETE е no-op, condition се игнорира. | `query/executor.nim:752` | В matched клона: провери `mergeMatchedCondition`, после honor-вай `mergeMatchedDelete` |
-| H9 | **WAL recovery чупи процеса при torn record** — recovery parser-ът вярва на `keyLen`/`valLen` (до ~4 GiB alloc) и `kind` (out-of-range enum → `CaseStmtError` Defect, не се catch-ва). | `storage/lsm.nim:704` | Bound lengths + валидирай `kind` преди use; дългосрочно per-record CRC32 |
-| H10 | **B-tree `remove` пише separator с грешна конвенция** — `splitChild` ползва left child max, `removeRec` пише right child min (`child.keys[0]`) → ключове стават ненамираеми при internal nodes (silent data loss). | `storage/btree.nim:377` | Separator = max ключ на left child, преизчислен след rebalance |
 
-### 🟡 MEDIUM (11)
+### 🟡 MEDIUM (7)
 
 | # | Проблем | Файл | Предложен fix |
 |---|---------|------|---------------|
-| M1 | **MVCC `write` трие от `activeTxns` по време на итерация** — `delete` proc-ът ползва collect-then-delete, но `write` трие inline (unsafe, пропуска timed-out txns). | `core/mvcc.nim:180` | Collect stale ids в seq, трий след loop-а |
 | M2 | **disttxn `connectWithTimeout` без SO_ERROR + uncaught RPC** — refused connect е "writable" → връща true; `sendDistTxnRpc` няма try/except → OSError wedge-ва 2PC състояние. (BUG-042 fix-нат в replication, не тук.) | `core/disttxn.nim:88` | `getsockopt(SO_ERROR)` + try/except около per-participant RPC |
-| M3 | **`checkpoint` leaking write lock при exception** — `acquireWrite` без try/finally; IOError от flush/rotate пропуска `releaseWrite` → постоянен hang. | `storage/lsm.nim:932` | try/finally около lock-а (и walLock) |
+| M7 | **Compaction unlink-ва input-ите преди output-ът да е loadable в каталога** — verifySSTable вече е преди unlink; остава catalog re-load ordering в caller. | `storage/compaction.nim` / LSM apply | Load/verify output в каталога ПРЕДИ unlink на input-ите |
-| M4 | **`flushUnsafe` празни memtable преди SSTable write** — при IOError на `writeSSTable` данните са загубени от memory (остават само в WAL, невидими за live reads). | `storage/lsm.nim:870` | Първо `writeSSTable`, после clear на memtable |
-| M5 | **Compaction пропуска empty-string ключа** — dedup sentinel `lastKey = ""` skip-ва ключ `""` → data loss при compact на празен ключ. | `storage/compaction.nim:130` | `haveLast` флаг вместо sentinel стойност |
-| M6 | **`rewriteLive` data-loss window** — `removeFile(wal.path)` преди `moveFile`; crash между тях губи unflushed записи. `rename(2)` и без това е атомен replace. | `storage/wal.nim:329` | Махни `removeFile`, остави атомния `moveFile` |
-| M7 | **Compaction unlink-ва input-ите преди output-ът да е loadable в каталога** — `applyCompactionResult` re-load-ва output в try/except (само warning); ако fail-не след unlink → загуба на ключове. | `storage/compaction.nim:150` | Load/verify output в каталога ПРЕДИ unlink на input-ите |
 | M8 | **`OFFSET n` без `LIMIT` връща 0 реда; negative `LIMIT` чупи** — `limitCount = 0` е sentinel и за "няма limit", и за "LIMIT 0"; `sourceRows[start..<endIdx]` с endIdx<start → IndexDefect. | `query/exec/lower.nim:411`, `plan_exec.nim:219` | Отделен sentinel (-1 = unlimited); clamp negative |
 | M9 | **Aggregate window функции връщат NULL** — `SUM/AVG/COUNT/MIN/MAX OVER (...)` попадат в `else` клона (`"\N"`); само ranking/lead/lag се handle-ват. | `query/exec/window.nim` | `of "sum","avg","count","min","max"` с `resolveFrameBounds` |
 | M10 | **WebSocket приема unmasked client frames** — RFC 6455 §5.1 изисква server да затвори връзката при unmasked client frame (cache-poisoning защита). | `core/websocket.nim:85` | Затвори връзката при `masked == false` |
 | M11 | **WebSocket без frame/message size limit → DoS** — `buf.add(chunk)` расте неограничено; няма 125-byte control-frame cap. | `core/websocket.nim:215` | Max frame/message size + control-frame cap |
-| M12 | **WebSocket SUBSCRIBE bypass-ва table-level auth** — всеки автентикиран клиент subscribe-ва към任意 таблица и получава всички insert/delete. | `core/websocket.nim:232` | Table read authorization при subscribe |
+| M12 | **WebSocket SUBSCRIBE bypass-ва table-level auth** — всеки автентикиран клиент subscribe-ва към произволна таблица и получава всички insert/delete. | `core/websocket.nim:232` | Table read authorization при subscribe |
 
-### 🟢 LOW (4)
+### 🟢 LOW (3)
 
 | # | Проблем | Файл | Предложен fix |
 |---|---------|------|---------------|
 | L1 | **SCRAM timing user enumeration** — unknown user връща веднага, known user прави urandom+HMAC/PBKDF2 работа → timing delta (BUG-049 fix-на съобщението, не timing-а). | `protocol/auth.nim:227` | Equivalent dummy work за unknown users |
 | L2 | **SCRAM channel-binding не се верифицира** — `c=` се приема verbatim; RFC 5802 изисква валидация. Не е exploitable днес (няма TLS-CB). | `protocol/auth.nim:259` | Enforce expected `c=` (напр. `biws`) |
-| L3 | **mmap read `offset + size` overflow** — `offset + size > region.size` с native int wrap-ва негативно при corrupt offset/size → OOB read. v3 SSTable-ите са CRC-защитени (reachable само през legacy v1/v2). | `storage/mmap.nim` | `offset > region.size - size` (без overflow) |
 | L4 | **NULL equality semantics** — `NULL = NULL` и `col = NULL` → true (string sentinel сравнение), не unknown/false. Системно за string-based value модела. | `query/exec/eval.nim:431` | Three-valued logic за NULL (по-голям рефакторинг) |
 
-### Хигиена
-
-- **Stray 94 KB компилиран ELF binary** в `src/barabadb/protocol/scram` — случайно commit-нат в source tree-то; да се премахне (+ `.gitignore`).
-
 ---
 
 ## Проверени и чисти (не са бъгове)

CHANGELOG.md

@@ -14,13 +14,28 @@ All notable changes to BaraDB are documented in this file.
 - **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`)
 
+### 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; 5 fixed in this batch, 23 tracked)
+- Deep audit report `BUG_AUDIT_2026-08.md` (~28 findings; 17 fixed across batches 1–2, ~12 tracked)
 
 ---
 

PLAN.md

@@ -163,7 +163,9 @@
 
 **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).
 
-**Остават (~23):** вж. `BUG_AUDIT_2026-08.md` — TLS peer verify, semi-sync partial-ack, COUNT(DISTINCT), UNION/INTERSECT/EXCEPT crash, MERGE THEN DELETE, WAL recovery crash, B-tree separator convention, MVCC delete-during-iteration, disttxn SO_ERROR, checkpoint lock leak, flushUnsafe data-loss, compaction empty-key, wal rewriteLive window, OFFSET-без-LIMIT, window агрегати, WebSocket (3), SCRAM (2), mmap overflow.
+**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.
+
+**Остават (~12):** вж. `BUG_AUDIT_2026-08.md` — TLS peer verify (H2), B-tree separator (H10, needs careful repro), disttxn SO_ERROR (M2), compaction catalog order (M7), OFFSET-без-LIMIT (M8), window агрегати (M9), WebSocket (M10–M12), SCRAM (L1–L2), NULL equality (L4).
 
 ---
 
@@ -179,7 +181,7 @@
 | **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) | 🔄 В процес — батч 1 завършен (5 поправки); `BUG_AUDIT_2026-08.md` |
+| **Сесия 13** — Stabilization & Deep Audit (2026-08) | 🔄 В процес — батч 1+2 (17 поправки); остават ~12; `BUG_AUDIT_2026-08.md` |
 
 ---
 

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/replication.nim

@@ -213,10 +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/query/exec/lower.nim

@@ -122,6 +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:

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)
@@ -292,49 +328,71 @@ 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/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):
+          let k = setOpRowKey(row, cols)
-            seen[valueToString(row["$value"])] = true
+          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)
@@ -759,7 +781,20 @@ 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 "$key" in tgtRow:
+          # Optional AND <condition> after WHEN MATCHED
+          var applyMatched = true
+          if stmt.mergeMatchedCondition != nil:
+            let condExpr = lowerExpr(stmt.mergeMatchedCondition)
+            applyMatched = valueToString(evalExpr(condExpr, rowWithTarget, ctx)) == "true"
+          if applyMatched and "$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:

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] = @[]

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
+  # Flush immutable memtable if present, otherwise flush current memtable.
-  var toFlush = db.immutableMem
+  # Do NOT clear the source memtable until the SSTable is written — an IOError
-  if toFlush.len == 0:
+  # mid-write must leave the data still visible to live reads (WAL still has it).
-    toFlush = db.memTable
+  var flushingImmutable = false
-    db.memTable = newMemTable(db.memMaxSize)
+  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:
@@ -930,7 +954,7 @@ 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)
@@ -946,10 +970,12 @@ proc checkpoint*(db: LSMTree) =
 
     # 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) =

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):
+    wal.stream = nil
-    removeFile(wal.path)
+  # 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

tests/bugfix_test.nim

@@ -625,3 +625,99 @@ suite "Query operator correctness — audit batch 1":
     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

tests/test_all.nim

@@ -1659,14 +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)
+    # No connected replicas → nothing to wait for; write succeeds
-    check not rm.isFullyAcked(lsn)  # still needs 1 more
+    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)
+    # ackLsn bookkeeping still clears pendingAcks at the required quorum
-    check rm.isFullyAcked(lsn)  # 2 acks received
+    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)