Resolved all 8 P0-P3 residual punch items from prior structural audit via wired-permanent solutions over one-shot fixes. (1) session-learnings-archive.md split into 4 shards (max 655 of 700 cap). (2)

Decision

Resolved all 8 P0-P3 residual punch items from prior structural audit via wired-permanent solutions over one-shot fixes. (1) session-learnings-archive.md split into 4 shards (max 655 of 700 cap). (2) Two 63-line feedback files compressed to 44 each. (3) Wiki entity catalog seeded at vault/wiki/index.md (cascade engine had no catalog → explained wiki/platforms/=0; future cascades self-populate). (4) Vault hot-cache manually refreshed. (5) vault-decision-outcome-sweeper.sh built + wired weekly Sat 19:30 UTC; initial run marked 112 stale decisions with Triage block for NOVA Sunday sweep. (6) 13 orphan self-maintenance crons wired with risk-tiered schedules (hook-registry-integrity, cron-dedup-guard, plugin-version-drift, openclaw-hygiene, vps-health-watchdog, oauth-convergence-check, cpu-spike-analyzer, etc.). (7) 6 P3a feedback files evaluated — verdict KEEP per format philosophy (within hard cap + cohesive content). (8) fabric+supabase MCP registration deferred to AJ via desktop tunnel.

Rationale

Structural fixes compound; one-shot fixes don’t. (a) Decision-outcome sweeper as cron preserves prediction-class decisions (pred-* prefix) and gives NOVA’s vault-hygiene-sweep something to act on weekly — vs batch-closing 128 entries blindly which would lose institutional signal. (b) Orphan-crons-wired honours Clause A of feedback_nothing_is_dead_weight (inactive = wiring gap, diagnose+wire, not retire). hook-registry-integrity cron specifically catches future cascade gaps like the openspace-quality-reporter deletion miss earlier this session — preventive control. (c) Wiki catalog seed lets cascade engine self-populate platform entities as 3+ mentions accumulate; manual creation would violate “Never modify wiki entity pages manually”. Premortem flagged risks: cpu-spike-analyzer auto-remediation could destructively restart containers on a false positive; mitigation = log + Telegram alert before any container action (script should be audited at day 7). claude-session-runaway-guard at 30-min cadence could kill a legitimate long Claude session; mitigation = grace threshold should be >2h CPU (verify script logic at day 7). Sweeper AUTOCLOSE_PREFIXES list is conservative — 0 auto-closures this run because oldest stale decisions used prefixes like bible-, agent-, xint- not in the list; future tuning needs broader prefix coverage. Bias check: anchoring on “13/13 wired” risks ignoring that wiring doesn’t mean correctly doing their job — 7-day post-deploy log review captures actual behaviour.

Alternatives Rejected

Outcome

Pending