CLAUDE.md: rollback-audit pass on detection layers (post 8-incident resolution)

SonicField · SonicField · commit f33bde699bc0 · 2026-04-23T21:44:06.000-07:00
Per supervisor 04:41:28Z (2) + theologian 04:43:15Z findings + supervisor 04:43:37Z accept-all-3 disposition. Adds rollback-audit subsection under Pre-Edit Working-Tree Integrity rationale. KEEP-all-3 with cost-acknowledgment: - 4-step Pre-Edit WT Integrity: KEEP (validated discipline regardless of false-actor framing) - mtime-checkpoint extension: KEEP (caught self-inflicted bug pre- commit at incident python#8) - Expanded pre-commit gate (compile + §3.5 BUILD MODE + per-bench): KEEP (validated; cost ~17min per Tier 8 commit acceptable) NEW cost-benefit rule: future detection-layer additions require explicit cost-benefit framing + cite the 8-incident false-actor lesson before authoring. trivial-push exemption (supervisor 00:39:20Z) remains in force. Sequencing slip: this was directed as part of cascade bundle per supervisor 04:41:28Z but landed separately. Acceptable per supervisor 04:43:37Z. Doc-only +2L (paragraph wrap; supervisor counted +6L logical). Authorization: theologian 04:43:15Z findings + supervisor 04:43:37Z GO.
diff --git a/CLAUDE.md b/CLAUDE.md
@@ -162,3 +162,5 @@ Rationale: 5+ prior incidents (D-1775810621, D-1775669703, D-1776414469, D-17764
 **Honest framing per pythia #105 2026-04-24:** This is an UNDIAGNOSED-RECURRING class. The 5+ prior incidents closed at the SYMPTOM layer (build break, gate fail, Alex directive) without confirmed root-cause attribution. This 4-step discipline codifies a reactive HALT — it does not provide automated detection. Detection still depends on the agent noticing mid-edit (which generalist did at 01:23Z without the discipline). Recurrence-prevention claim does not exceed what this text mechanizes; automated detection (file-watcher / git-hook) feasibility is fixup PIR territory and remains open.
 
 **Update 2026-04-24T04:39:18Z (post Tier 8 Phase A landing):** The 8-incident class was RESOLVED by root-cause attribution at b83f0840fb (push 31): the apparent 'external reverter' was our own `scripts/w45_section_3_5_derivation_drift.sh` restore_files() trap using `git checkout HEAD --` on TOUCHED_FILES, destroying pre-existing unstaged user modifications. Fix: snapshot-based restore preserving pre-script state. mtime-checkpoint discipline VINDICATED — caught self-inflicted infrastructure bug pre-commit (testkeeper 03:53:05Z). Lesson: 'external actor' hypotheses need a `grep -lE 'git checkout HEAD --' scripts/` first-step BEFORE detection-layer escalation. Per feedback_falsifier_restore_clobbers_unstaged.md (librarian 04:28:11Z institutional record).
+
+**Rollback-audit 2026-04-24T04:41:28Z (per pythia #109 + supervisor):** The 4-step Pre-Edit WT Integrity discipline + mtime-checkpoint extension + expanded pre-commit gate were authored under FALSE-ACTOR framing (assumed external reverter). Root cause was internal infrastructure. Rollback decision: KEEP all three layers (each is independently validated discipline). Latency cost acknowledgment: expanded pre-commit gate adds ~17min per Tier 8 / multi-file builder*.{cpp,h,c} commit (compile ~80s + §3.5 BUILD MODE ~6min + per-bench gate ~10min). Cost justified for that commit-class; supervisor 00:39:20Z trivial-push exemption (script-only, doc-only) remains in force — DO NOT extend full gate to trivial pushes. mtime-checkpoint cost is per-write (~5s per touched file); acceptable for any multi-file edit. Future detection-layer additions: require explicit cost-benefit framing + cite the 8-incident false-actor lesson before authoring.
