fix(learn): canonical stateRoot resolution + scope-routed guideline synthesis

[cmschuetz]

Why

Investigating "is /strapped:learn broken? feedback isn't being collected" surfaced that feedback was being collected (10 unsynthesized critiques across 5 runs) — the failure was a false negative: learn was the only strapped skill that never resolved stateRoot via state.mjs, so any mis-resolution glob'd to zero and reported "no critiques, stop" — indistinguishable from genuinely none.

A second, deeper issue: learn dumped every synthesized guideline into one CLAUDE.md, whose load scope is only "working in the plugin repo." So harness-behavior rules learned from runs never fired for runs against other target repos — the majority of intended use.

Heads-up (from the requester): this likely won't merge as-is — it's a reference point for the direction.

What

Resolver gap (the concrete bug)

• New state.mjs runroot — slug-less { stateRoot, runRoot } resolution that hard-fails on an unresolvable/relative anchor, so an empty cross-run glob is unambiguously "no runs" and never a silently-wrong root. Documented in conventions.md; +3 CLI tests.
• learn Step 1 now resolves via runroot, echoes the resolved root + match count, and treats a resolution error as a loud stop.

Scope routing (the design fix)

• learn Steps 2–5 now classify each cluster harness-behavior vs repo-development and route it:

  • harness rules → most specific stage prompt (src/workflows/strapped-run/stages/*.ts, review-loop.ts) / SKILL.md step / conventions.md
  • repo-dev rules → the pertaining repo's CLAUDE.md (resolved from the run's repos map), not necessarily the plugin's.

• The 4 generalizable guidelines synthesized this pass were relocated out of CLAUDE.md into their targeted prompts so they reach every run:

  Guideline	New home
  Verify plan claims vs origin/main (fetch first)	plan.ts planner
  Add a real test suite as the validation gate	plan.ts planner
  Review design decisions vs the ask's intent	review-loop.ts completeness lens
  No grep-guard prose regression tests	implement.ts implementer
  Reconcile plan vs current main before implementing	implement/SKILL.md pre-flight

• CLAUDE.md keeps only genuinely repo-dev rules (new-repo/breaking-changes-fine, no-machinery-without-signal) + a pointer to where harness guidance lives.

• Plugin 0.6.0 → 0.7.0.

Not addressed (flagged)

Critique writing remains a pure model-followed instruction with no hook/script enforcement — autonomous /strapped:run flows skip it silently. Left as a follow-up.

Test

bun run typecheck && bun run lint && bun test → 109 pass (+3). Generated strapped-run.js rebuilt and in sync; all 4 relocated snippets verified present.