Keep how-to replies focused, complete and appropriately routed

[aryanorastar]

Customer failures prevented

Technical topic words and navigation/confusion or hypothetical clauses caused unnecessary handoffs; long answers were cut mid-item; multipart questions lost complementary official pages; replacement intake requested tracking instead of damage evidence.

Changes

• Keep APP/DESKTOP/FIRMWARE evidence lanes, but escalate only reported failures. Navigation predicates are removed without deleting independent incident evidence. Conditional clauses are excluded while observed past incidents and independent failing consequences remain reports. Protected human, money, privacy/account and data-loss routes are unchanged. No complete review/evaluation sentence was copied into code.
• Answer prompt carries an approximately 1,500-character budget and brief supported parts in requested order. Overflow presentation retains complete bullets/steps/wrapped items or prose sentences, complete source URLs and protective data-loss warnings.
• Preserve neutral sub-question/search-query pairs. More than three parts may retrieve four complementary pages within eight chunks and the unchanged 9,000-character evidence cap. Production and evaluation share the coverage policy.
• Physical damaged/replacement intake asks for photos/video, private order email and delivery date. Unrelated account/receipt and parcel intake remain unchanged.
• Merge current main into the branch without rewriting history; document current release gates.

Verification at 251d20e

• npm test: 707/707 pass before push; offline customer check 8/8. Hosted Node 20/24 and database CI pass.
• Latest routing correction has six own English/Spanish how-tos, six own incident reports and boundary/negative regressions. Seven focused tests failed before the correction; all 16 new tests pass after. An additional unpunctuated conditional-consequence fault was reproduced and preserved.
• The earlier candidate run stopped after eight saved rows with no recorded row error/timeout. Exit/signal metadata was not retained; no matching sleep transition or crash report was found. Exact crash-versus-kill cause is unknown, not asserted as host sleep. Private diagnosis evidence is retained outside Git.
• Five new candidate cases ran serially without errors, each under 90 seconds: 66.197, 79.940, 34.697, 56.338 and 53.643 seconds.
• Fresh same-day comparison completed on frozen runtime 251d20e, against main b9eeb16: 162 valid rows per arm (81 × 2, concurrency 3). Main 02:25:04–03:09:18 UTC; candidate 03:09:18–03:53:25 UTC; inter-arm gap 81 ms. Both recorded normal scoring exit 1, no termination signal or logged execution error; supervisor exit 0. Shared environment/harness/lockfile and clean commits were guarded. Temporary awake assertions were verified. Questions were not inspected or used for tuning. Live source/provider variation and non-isolated host activity remain limitations.
• Mutually exclusive pass-both/flaky/fail-both: main 71/6/4, candidate 70/5/6. Total wall p50/p95: main 50.601/79.179 s, candidate 49.957/80.067 s. Numerical gates pass: 70 ≥ 69, p50 delta −0.644 s, p95 delta +0.888 s.
• Quality approval is still held. Structural unwanted-handoff flags increased 12 → 15; missing-handoff flags are 2 → 2. These counters do not replace independent factual/how-to/protected-escalation grading. Source-label duplicates: 1 → 0; no ending-ellipsis/ellipsis-before-source pattern found, but semantic clipping still needs review. Six second reviews per arm; 18/162 matching replies byte-identical. Both full JSONs, stage percentiles, per-case escalation deltas, interrupted-run diagnosis and provenance are retained privately outside Git. No raw replies are published here and no implementation tuning from protected results is performed.

Release hold

Draft, not approved to merge or deploy. Reports 2/3 passed; production stays 4a8e93d5 / main b9eeb16, rollback 6ddd1673. Remaining order is #42 after its gate/review → closing-card localization #43 after review → #28 after its private-destination preflight/review → #29 → #30. #43 separately passes 668 tests and hosted CI. #28's 750-test preparation is local only and stopped on actual untrusted_role_can_view; no permission or allowlist changes, push, merge or deployment. One deploy at a time, 24 clean hours unless explicitly waived. No Railway or credential changes. Diagnostics remain parked and the deferred credential-rotation risk remains unresolved.