2026-08-08 — KB integrity fix: capture runaway closed, KB dedup'd 7,124→102 real facts, EAT_ALL 21/21 green

KB incident + recovery + hardening (2026-08-08, JEEVES lane).

The incident (honest record)

A --refine timeout corrupted the canonical KB. Root causes found and fixed:

1. zsh preexec capture hook logged every terminal command (grew captures to 787MB).
2. refine/gNN readers scanned the whole capture each run (227s+ stalls).
3. Feedback loop: refine re-wrote events it read — KB ballooned 85 → 328K entries.

The fixes (all committed)

• _build_terminal_kb_entry — no-logging KB builder (feedback loop closed)
• REFINE_WINDOW=500 — bounded window on all capture readers
• normalized-question dedup — KB 7,124 → 102 real facts (~99% was redundancy)
• compact json.dumps() (no indent=2) — KB loads in 0.00s
• capture .zshrc hook size-capped at 10MB/day
• EAT phase_9i: 227s → 0.4s

Verified green

• E2E batch: 11/11 PASS
• EAT selftest: 13/13 · flywheel: 15/15
• EAT_ALL: 21/21 phases, 0 failed, 0 skipped (first fully-green since the incident)
• KB: 102 entries, 138KB, all 6 apices HTTP 200

Moat intact

The KB is now a clean, growing knowledge base (genuinely-new facts only). The 6-layer moat (per the estate canon) is healthy.