[aryanorastar]

Customer failure being prevented

Before a memory pilot, separate passing storage and context-reader tests do not prove that saved SQL pointers actually reach only the right customer's live posts. A disconnected reader, stale context or forgetting that does nothing could escape that coverage.

Test-only change

• Add one synthetic rehearsal to the existing guarded local/Postgres CI verifier. It composes the production SQL memory adapter, service and live-context reader; no production modules, schema, prompts, dependencies or settings change.
• Eight scenario groups cover two-post scope/role boundaries, uncached edits, reader recreation, changed-owner rejection, deleted-message/outcome invalidation, context-off reads, capture-off forum erasure, durable opt-out/expiry and text-free rows before erasure.
• Fake post/message objects make no Discord or provider calls. The verifier's existing local test-database restriction remains; PGlite is only a serialized local alternative, not concurrency proof.
• Document the distinction between this rehearsal and enabled generated answers, a full bot restart, the live pilot and remaining privacy/data-use gates.

Evidence

• Local full suite 839/839, offline 8/8, isolated SQL verifier including all eight rehearsal groups pass.
• A separate private, in-memory sensitivity run passes the real baseline and rejects five deliberate faults: disconnected reader, missing requester-owner guard, cached live reads, stale deleted outcome, and successful-but-no-op forgetting. Runtime repository files are not modified by those probes. This is test sensitivity, not five observed production bugs or runtime repairs.
• Hosted Node 20/24/Postgres pass at exact head c2953f27e7f3afc48b1ed5f6be5e00f823153e86 (run 39546418506). The database job's log confirms all eight SQL/live-reader rehearsal groups ran against its real PostgreSQL 16 service, not PGlite.
• October 23 fresh local verification on macOS, Node v26.7.0: full suite 839/839, offline 8/8, and all eight SQL/reader scenario groups pass again. The real baseline passes; five independent in-memory faults are all rejected. No repository runtime file is mutated, external service called or production database touched. Local PGlite 0.5.10 remains serialized; hosted PostgreSQL supplies the separate real-server integration evidence.
• One user-authorized independent review inspected the exact head, test assertions, isolation guards and hosted execution log. No actionable findings. The SQL adapter → memory service → live-context reader harness is the appropriate running-environment end-to-end check for this test-only change. This is an independent code/evidence review, not a fabricated human GitHub approval.

Reproduce safely

On an installed checkout of this head: npm test, EVAL_NO_MODEL=1 node scripts/customer-eval.js, then node scripts/verify-support-db.js with the verifier's dedicated local test database, or node scripts/verify-support-db.js --pglite /absolute/path/to/@electric-sql/pglite. The database guard rejects non-test targets. Expected: 839 tests, eight offline checks, the verifier's existing SQL checks and the eight named joined-reader groups pass. Never use production credentials or a production database. The separate sensitivity probe is private verification tooling, not a dependency of this PR.

Release boundary

The affected test-only flow is verified and independently reviewed; ready for maintainer review, not merged or deployed. Memory/context/fix-inbox remain off. C1–C4, enabled generated answers, full bot restart/pilot, processor/backup/security/notice approval and actual memory enablement remain separate open gates. No new paid comparison or protected input read. Existing comparison holds and earlier release decisions are unchanged. Canonical current product/release status remains in documentation PR #69; this branch is based on main and its older release checkpoints are historical. No delivery-score increase.