Distillation

Date: Sep 16, 2026

• 🔴 (12:02) User’s regression-test discipline requires every adversarial-review defect to land with a deterministic regression that fails on the base branch, passes on the fix, exercises the real precondition rather than artificial state, and fails if the tested guard is deleted.
• 🔴 (12:02) User directive: “Never trust the fix without seeing it fail.” Every bug fix must be driven failing-first: write the regression first, confirm failure on the base branch, then confirm it passes with the fix.
• 🔴 (12:02) User’s adversarial state setup requires: 1. state already present before the operation that should handle it; 2. delete-then-recreate sequences across lifecycle boundaries; 3. modify-while-off sequences where capture triggers cannot observe the change.
• 🔴 (12:02) User’s fan-out registry contract batteries are: 1. SYNCED_TABLES → packages/core/test/sync-registry-contract.test.ts, covering table shape, pull-only/no-capture invariant, and composite primary-key correctness; 2. MIGRATIONS → db.test.ts schema-version assertion plus migration-specific tests, including one toBe(MIGRATIONS.length) assertion; 3. AGENTS → agents.test.ts, requiring every agent to expose name, displayName, binary, and detect(), with binary a constant placeholder string.
• 🔴 (12:02) User’s standing bug-class batteries are: 1. registry contracts for SYNCED_TABLES; 2. per-operation sync invariants in packages/core/src/sync-data.ts via assertSyncInvariants, enforcing no pull-only outbox, profiles ≤ 1, and every state row referencing a registered table; 3. sequence properties in packages/gateway/test/sync.property.test.ts, covering invariants after every operation, prune-floor liveness, eventual convergence, no ping-pong, and currentTier correctness.
• 🔴 (12:02) User’s review heuristics identify skip/early-return branches as the highest-risk surface; every skip branch needs a test that makes it fire wrongly and proves it does not skip required work.
• 🔴 (12:02) User’s review workflow requires two distinct passes: 1. adversarial/correctness review covering enumerated break states, apply-then-revert mutants, property-test stability ×10, hash-equivalence proofs, frozen-array mutation safety, and stale-cache analysis; 2. security/pentest review covering injection vectors, credential leaks, protocol-level attacks, and privilege escalation.
• 🔴 (12:02) User’s reviewers must run pnpm test, pnpm run typecheck, pnpm run lint, and pnpm run format:check; inspect exit codes because blank output tails can hide errors; run property tests ×10; hash-verify that the working tree remains unmodified; and provide per-point PASS / FAIL / CONCERN / MUST-FIX verdicts with file:line evidence plus an overall MERGE / DO-NOT-MERGE verdict.
• 🔴 (12:02) User’s correctness-critical paths—sync engine, hashing, lifecycle, and migrations—require adversarial review regardless of change size; mechanical or test-only PRs may skip it at reviewer discretion.
• 🔴 (12:02) User stated frozen shared arrays are preferred for memoized results shared by multiple callers, so accidental mutation throws TypeError in strict mode rather than silently corrupting the cache.
• 🔴 (12:02) User stated db().query monkeypatches are shadowed by the tracing Proxy; SQL-counting tests must use the log.registerSink({ withDbSpan }) seam.
• 🟡 (12:04) Assistant added a live-cgroup-move regression to packages/core/test/cgroup-memory.test.ts; it calls the public headroom reader twice while /proc/self/cgroup membership changes, so an implementation caching the old cgroup path would return the stale budget and fail for stale admission.
• 🟡 (12:06) The focused packages/core/test/cgroup-memory.test.ts run passed: 1 test file passed, all 14 tests passed, duration 1.51s (transform 932ms, setup 1.02s, import 19ms, tests 15ms, environment 0ms).
• 🟡 (12:06) Assistant concluded the live-move regression passes because the resolver re-reads membership rather than caching it, closing the fourth identified cgroup defect.
• 🟡 (12:06) Assistant began tracing two remaining admission failures through composed call paths: 1. unknown or conflicting primary-worker headroom; 2. burst pool growth that can reserve the same available bytes more than once.
• 🔴 (12:06) User’s adaptive embedding cap never exceeds MODEL_MAX_TOKENS.
• 🔴 (12:06) User’s memory-limit policy is fail-stable when free memory never improves: it must never re-probe.
• 🔴 (12:06) User’s persisted embedding-cap policy must never re-probe upward to a cap beyond what the WASM heap supports.
• 🔴 (12:06) User’s embedding pool size is never below 1; the primary worker always runs.
• 🔴 (12:06) In packages/core/src/embedding/local.ts, availableMemoryHeadroom() combines constrainedMemoryLimit(), Linux readLinuxCgroupMemoryHeadroomBytes(), and when available process.availableMemory() through resolveMemoryHeadroom(); Bun omits process.availableMemory() because its implementation currently aliases host freemem() inside constrained cgroups.
• 🔴 (12:06) In packages/core/src/embedding/local.ts, containerFreeBytes() clamps host freemem() through clampFreeToContainerLimit(raw, constrainedMemoryLimit(), availableMemoryHeadroom()).
• 🔴 (12:07) In packages/core/src/embedding/pool.ts, the primary worker is spawned whenever this.slots.length === 0, with its own OOM backoff rather than pool sizing intended to protect constrained hosts.
• 🔴 (12:07) In packages/core/src/embedding/pool.ts, pool growth currently requires at least one healthy slot, all eligible workers busy, total active plus retired workers below this.ceiling, and liveFreemem() at least (number of unhealthy slots + 1) * PER_WORKER_MEM_BUDGET_BYTES.
• 🔴 (12:07) In packages/core/src/embedding/pool.ts, liveFreemem() returns 0 when constrainedMemoryLimit() > 0 but availableMemoryHeadroom() is undefined; otherwise it calls clampFreeToContainerLimit(raw, constrained, headroom).
• 🔴 (12:07) In packages/core/src/embedding/pool.ts, each new LocalProvider receives the pool ceiling as its memory divisor, so each worker sizes its token cap from free / ceiling; the intended invariant is that provisioned workers collectively remain within one EMBED_MEM_FRACTION share of free memory rather than each independently claiming half.