Substrate governance: skill bloat from premature codification — codify-deferral discipline

Context

Operator framing earlier in 2026-05-18 session (#11598 META): observable pattern of new lessons immediately codifying as skill substrate vs. accumulating into single bigger lesson. Symptoms surfaced as recurring "skill bloat" friction.

V-B-A this turn against last 200 merged PRs: at least 43 gate-additions in 9 days (feat(ci) / feat(hygiene) / feat(agents) / feat(skills) / feat(agentos) with gate/guard/invariant keywords). Many are skill-substrate accretions on top of skill-substrate accretions.

Friction

• Each lesson immediately becomes its own skill / its own SKILL.md section / its own §X subsection
• Skill substrate accretes faster than it consolidates (no Lessons → Skill aggregation cadence)
• Loaded-byte cost compounds; skill discovery becomes harder; routing rules become more numerous
• Distinct from monkey-patching (symptom-vs-root); this is the granularity failure: right diagnosis, wrong codification scale
• Example shapes from recent merged PRs (see V-B-A): feat(ci): CI lint for agent PR review bodies + feat(ci): enforce agent PR body template adherence + feat(ci): two-layer body-shape validation + feat(github-workflow/mcp): mechanical body-shape validation — four separate substrate-amendments for what is one underlying lesson (agents produce malformed bodies)

Scope

Codify a codify-deferral discipline:

1. Accumulation period — new lesson held in scratch substrate (private memory / Discussion comment) for K weeks before promotion to skill substrate
2. Promotion criteria — promotion requires N independent recurrences across multiple agents/sessions, not single-incident reactive promotion
3. Consolidation cadence — quarterly (or per-N-skill-additions) audit that groups related lessons into single substrate section vs. parallel sections
4. Anti-pattern detector — same lesson scope codified in M places = consolidation trigger

Acceptance criteria (stub)

• Codify-deferral discipline named in ADR or workflow doc (likely ADR 0008 SKILL anatomy amendment, since it owns skill-shape rules)
• Accumulation period K + promotion criteria N defined (empirical-anchored, not arbitrary)
• Consolidation cadence + detector signature
• First-cycle dry-run identifying N current-substrate consolidation candidates
• Cross-reference with sibling friction tickets (monkey-patch, substrate-bandwidth)

Empirical anchor

• Operator framing 2026-05-18 session (Discussion #11598)
• V-B-A: 43 gate-additions / 195 META in 9 days; body-shape-validation cluster (4 separate PRs for one lesson) as concrete consolidation case
• Companion: ADR 0007 compaction taxonomy is the substrate-side mechanism; this ticket addresses the before-substrate side (don't codify yet)

Cross-refs

• Refs #11598 META — friction item NOT previously captured as ticket (operator-correction)
• Related substrate: ADR 0008 SKILL anatomy + authoring contract, ADR 0007 compaction taxonomy, AGENTS.md §friction_to_gold (which can be misread to mean "codify immediately" — this ticket clarifies)
• Sibling tickets: monkey-patch vs root-cause; substrate-bandwidth-saturation

Priority

Team to triage. Related to substrate-bandwidth ticket but addresses a distinct mechanism (granularity vs. surface-symptom).