[aryanorastar]

Customer failure prevented

Ordinary support replies should keep the next step in the current thread. Delivered handoffs must not simultaneously tell the customer to email elsewhere. A failed staff delivery must retain the application's honest email fallback.

Draft: CI publication only. No merge/deploy while the reliability-release smokes are pending. Current main 6b6d97d1168da34be012bb7936e9f1f02da82966 integrated with merge commit dc79d80; only document conflicts needed resolution. No new PR, prompt/provider call, source budget, memory switch or production change.

Required review correction

• Detect support-address clauses independently of language, except the existing secret-sharing warning exemption. Keep existing English address-less patterns. Address-less non-English redirects remain explicitly outside scope in docs/NO_EMAIL_REDIRECT.md.
• Reuse shared source-label names/separators for localized/plural/full-width Source preservation and citation-only emptiness. Retained intake details cannot restore a stripped address.
• Keep the application-owned failed-staff-delivery email fallback: append it after cleanup, never claim receipt on failure.
• Preserve the original ordering: ordinary cleanup runs only after no escalation was selected; failed-self-service detection adds a handoff, never suppresses it. Redirect-only drafts still take a real handoff.

Verification

• 943/943 npm tests, 8/8 offline customer eval, isolated SQL pass, including current recovery, reference-memory reader and knowledge-retention checks. Fresh npm test before push.
• 25 new synthetic regressions, not copied from evaluation replies: French/German/Japanese/Portuguese ordinary and delivered paths in helpers plus actual help-thread handlers with deterministic synthetic adapters; source/secret-warning/failed-delivery protection; intake-tail safety; seven localized/plural Source headings and source-only emptiness.
• Same final helper/handler tests with pre-repair dc79d80 utilities injected only in memory: 160 pass, 25 fail. Repaired: 185/185 pass. Existing actual failed-delivery handler tests also pass.
• Independent reviewer reproduced the Source-label issue; seven regressions failed before its repair, then passed. Final narrow independent re-review: PASS, 185 affected checks, no remaining actionable finding. No provider/Discord/customer/production access by reviewer.
• Exact head f940c2766d978fc48c384a053c7c131b56707107: hosted Node20, Node24 and real PostgreSQL pass, run 39981758019. Repair-head run39981028525 also passed; the final commit only records published checks in the master plan. Local SQL is not substituted for multi-connection CI.
• One later local status-doc check had942/943 passing: the unchanged inactivity-only loopback demonstration aborted its stream while two local suites ran. No code/test change hid it. All six deadline tests pass separately; the fresh full serial npm test -- --test-concurrency=1 passes943/943 before the final push. The fixture depends on real40ms ticks under a100ms timeout; host scheduling is plausible, not proven. No production regression or fixed-flake claim follows.

Frozen comparison provenance and grade

Original main arm: 5bafe93f305725db1afc4b2f3e84bb3acc0e0dd2.
Original candidate arm: f2781ebaad8bbcb14c57db8f5cd390ae59a43d1d; its parent is the exact main above. Thus both arms used the same base. Harness/judge/lockfile matched; each completed81×2 at concurrency3, zero invalid rows. Result files are unchanged; this is not a new comparison of the integrated cleanup.

Independent grade: protected-handoff gate PASS; measured p50 30.255→32.455s (+2.200), with an explicit reviewer exception for the extra0.200s. p95 47.351→52.850s; pass-both73→72. Extra/missing flags and the unresolved own-key web-app answer miss remain recorded, not relabeled as fixed. Review finds the handoff-flag differences arise in generation, not escalation suppression by this diff.

The cited official backend storage page was read successfully via direct HTTPS on October23. It supports source/device constraints on automatic merging and retention of original recordings, not a universal promise that every source always appears separately. Today's page is not proof of its exact October21 contents. Grounding nuance and the own-key answer-quality miss remain A2/E2 work; no case-specific patch.

Queue and release limits

#62 stays open for narration cleanup and per-stage evidence. Reviewer confirms #77 overlaps: leave it draft untouched until #67 merges, then integrate main and drop duplicated redirects. Do not touch #78 until the controlled drain smoke is reviewed. #74 was published for CI only. No merge/deploy or memory enablement from these checks; product progress remains47/100, not accuracy.