Recover Stranded Signals Work, Reduce Smith's Instruction Load

Timeline

1. 1

   Start

   Stale main-checkout ownership claim found on session start

   Prior session ended without releasing; force-released after human confirmation

2. 2

   Recovery

   Signals Cockpit ranking fix verified and landed

   tsc, full suite, RSC boundary check, real production build — then found findings.ts was dead code, not wired to the UI as the prior handoff claimed

3. 3

   Redirect

   New Charge arrived: reduce Smith's default instruction load

   Zeb flagged tokens being consumed faster than normal mid-session

4. 4

   Investigation

   Baseline measured, Council-reviewed plan narrowed twice

   Model switched to Opus; scope cut from a full rewrite to a single bounded Archived Traps extraction after a subagent report was found unreliable

5. 5

   Execution

   Archived Traps moved, guard hardened, fixture-tested

   Real defect (silent-zero on missing archive) found and fixed as part of the move, not just documented

6. 6

   Close

   Council review caught 2 real defects in the just-shipped commit

   Duplicate rows across Active/Archived, and an incomplete before/after measurement — both fixed same session

What shipped

• Committed Signals Cockpit ranking fix (04fe6072): occurrencesLast10 replaces lifetime painScore as the sort key for "Needs a Charge", RECENCY_WINDOW moved to categorize-spin.ts to fix a real RSC-boundary production build error, Owner Loop Lifecycle CATEGORY_COHERENCE entry (11 spins, corrected twice under review), UI Layout classifier-contamination fix

• Corrected the Signals Cockpit handoff Charge (66d00d63) to state plainly that findings.ts is dead code and the two-section UI split was never built, replacing a prior session's inaccurate claim that it only needed verification

• Moved Archived Traps out of ignition-core.md into forgekit-os/archived-traps.md (e3a5c0aa), closing a real false claim ("archived traps are not loaded by default" while physically loaded every session)

• Reviewed all 13 archived rows for live operational content before moving them; promoted 2 (UTC-drift dates rule, Clerk-auth timebox stop condition) back to Active after grep confirmed zero coverage elsewhere; demoted 2 verified duplicates of eagerly-loaded gate checklists to make room

• Fixed a real latent defect in engine-guard.mjs: countTableRows() returned 0 for both a missing file and a missing header, which would have silently disabled the archived-row cap forever if the archive file were ever lost or malformed. Missing/malformed now hard-fails.

• Fixture-tested all 4 archive failure paths (missing, malformed, over-cap, warn-at-18) plus exit codes against temp copies, without touching the real archive

• Individually rebaselined the 3 hard-rule-baseline.json entries affected by the line shift, verified by phrase/order/text rather than regenerated wholesale

• Added 'Where an instruction belongs' maintenance guidance to HOW_TO_WORK_WITH_ZEB.md

• Found and fixed, under the same review pass, a real defect in the just-shipped e3a5c0aa commit: the Dates and Auth/Clerk-timebox rows existed in BOTH the Active table and the archive after being promoted — the commit's own message had claimed 'no rows lost or duplicated'. Fixed in d610d7c8; row totals now reconcile exactly (20 active + 13 archived = 33, matching the starting count)

• Corrected the reported byte-reduction figure in the same fix: the original 3,161-byte claim counted only the .claude/rules/ surface and omitted +1,139 bytes added to HOW_TO_WORK_WITH_ZEB.md (a mandatory orientation read). True net reduction is 1,938 bytes (~0.6%), not 3,161 (~1.1%)

• Committed findings.ts (9068ecf6) explicitly labeled as known-unwired, rather than left stranded in the working tree or silently implied to be shipped

• Restarted Alloy dev server cleanly on port 3911 (.next wiped, confirmed serving via curl)