session 2026-09-03: the index had become the record — breach fixed, Active Session rotated

Session record + MEMORY.md rotation (prior Active Session demoted verbatim to
MEMORY-reference.md, 9 of 9 non-blank lines conserved) + 4 KG rows.

MEMORY.md 23,714 -> 20,567 bytes; headroom now 4,419. The trim itself landed in
2171d91; this is the wrap half.

Recorded for the next wake: the wake digest's --selftest now fails on every run.
Its end-to-end control samples the 13 most recent transcripts and all 13 are
Tarbuckle mumbles (~66 KB, one human turn each) — PENDING-178's mumble pollution
displacing real sessions out of a second instrument. Not filed; the moratorium is
undecided and this belongs with -178 when that is ruled.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01X3L79vgAnt1x2kxvf23Qt7

claude/memory/knowledge-graph.jsonl

@@ -769,3 +769,7 @@
 {"subject": "an instrument that names no defect in advance", "predicate": "prevention", "object": "FOUND WHAT NOBODY PREDICTED, WHICH IS THE ONLY EVIDENCE AN INSTRUMENT IS NOT CONFIRMING ITS DESIGNER. verify_typography.py was built converter-agnostic by construction; on its first run it found the defect it was built for (curly_single 4608 -> 4, word-boundary damage) AND one nobody had looked for (ligatures 2013 -> 0, benign on reading). The division held: the instrument reported the vanishing, a human judged it harmless in the artifact.", "valid_from": "2026-09-01", "valid_to": null, "confidence": 1.0, "source_file": "session-2026-09-01-the-gate-that-should-have-existed.md", "extracted_at": "2026-09-01"}
 {"subject": "constraint-6-differently-biased-checkers", "predicate": "evidence-for", "object": "ALL TWELVE OF THE DAY'S EXECUTOR ERRORS AND THREE SELF-REPORTED JURIST ERRORS WERE CAUGHT ACROSS PARTY LINES; NOBODY CAUGHT THEIR OWN. Steward caught the executor's dates and arithmetic; jurist caught its numbers and an overreach; executor caught the jurist's unread rows and a read surface (governance-mcp.py --selftest) failing WHILE the jurist ruled from it. The misses did not correlate — which is the falsifiable claim Constraint 6 makes and asks to have watched.", "valid_from": "2026-09-01", "valid_to": null, "confidence": "high", "source_file": "session-2026-09-01-the-gate-that-should-have-existed.md", "extracted_at": "2026-09-01"}
 {"subject": "PENDING-134", "predicate": "ruled-by", "object": "REVIEWED-121, 2026-08-14. MEMORY.md carried 'NEXT: rule PENDING-134' for 18 days after it was ruled and aimed the steward at the wrong target; the live blocker was PENDING-137, ruled as REVIEWED-133 on 2026-09-01.", "valid_from": "2026-08-14", "valid_to": null, "confidence": 1.0, "source_file": "session-2026-09-01-the-gate-that-should-have-existed.md", "extracted_at": "2026-09-01"}
+{"subject": "claude-code", "predicate": "drift-pattern", "object": "MY OWN CONSERVATION CHECKER WAS CASE-SENSITIVE AND REPORTED A FALSE LOSS. Verifying the MEMORY.md relocation, the checker claimed 13 tokens had been lost; three were sitting in the target file in capitals ('PHOTOGRAPHED PAGES OF PHYSICAL BOOKS'). Re-run case-insensitively the genuine gap was 2 facts. An instrument that reports false losses is dangerous in BOTH directions: it trains the operator to discount it, and the day it reports a real loss it will be discounted too. No positive control had been written for it before first execution.", "valid_from": "2026-09-03", "valid_to": null, "confidence": 1.0, "source_file": "session-2026-09-03-the-index-had-become-the-record.md", "extracted_at": "2026-09-03"}
+{"subject": "claude-code", "predicate": "drift-pattern-good-direction", "object": "DIAGNOSED BEFORE PRUNING, AND THE DIAGNOSIS INVERTED THE ACTION. Asked to fix MEMORY.md's load-budget breach, the obvious move was to cut the longest entries. A per-line byte map plus a duplication check found instead that the index had become the SOLE CUSTODIAN of live state for three workstreams, and that the longest entries in 'Rules that fire silently' are long BY DESIGN. Pruning would have deleted three undocumented workstreams and removed the catches that fire when I would not know to look them up.", "valid_from": "2026-09-03", "valid_to": null, "confidence": 1.0, "source_file": "session-2026-09-03-the-index-had-become-the-record.md", "extracted_at": "2026-09-03"}
+{"subject": "the discipline 'prove the instrument before trusting a clean line'", "predicate": "prevention", "object": "CAUGHT A PERMANENTLY-FAILING CONTROL BEHIND A CLEAN REPORT, AND FOUND PENDING-178 AT A SECOND SITE. wake-digest.py reported 407 pointers / 0 dead; its --selftest reported 86 of 87 with overall FAIL. The failing control samples the 13 most recent transcripts and ALL 13 are Tarbuckle mumbles (~66 KB, one human turn each) — the mumble pollution filed as PENDING-178 against the ladder counter has also displaced real sessions out of the digest's own end-to-end control, which will now fail on every run. A lesson banked about instruments stopped a different failure class: alarm-decay, not a false alarm.", "valid_from": "2026-09-03", "valid_to": null, "confidence": 1.0, "source_file": "session-2026-09-03-the-index-had-become-the-record.md", "extracted_at": "2026-09-03"}
+{"subject": "MEMORY.md", "predicate": "failure-mode", "object": "AN INDEX BECOMES THE RECORD WHEN ITS TRACKER GOES QUIET. Observed 2026-09-03: the studium tracker's chronological log stopped 2026-08-13 while the MEMORY.md one-line pointer carried engine state to 2026-09-01; the Fool's only link was a SEALED seed with nowhere to append; L1's replay mechanism and completion criterion existed in the index alone. The tell is a tracker going quiet while its one-line pointer grows. The rotation at wrap handles the Active Session block; nothing handles slow growth in the tracker list, so the mechanism that produced this breach is unfixed.", "valid_from": "2026-09-03", "valid_to": null, "confidence": 1.0, "source_file": "session-2026-09-03-the-index-had-become-the-record.md", "extracted_at": "2026-09-03"}