rigor-reviewer: Seal Level 3 remediation loop (triage + consent-gated auto-fix)

[ShokuhouMisak1]

What

Closes the ARA pipeline loop compile → record → review → fix by adding a Seal Level 3
remediation capability to rigor-reviewer, kept self-contained (no new standalone skill).

Changes (only skills/rigor-reviewer/)

• Per-finding triage (review_version → 3.1.0): every Level 2 finding now carries
  fix_class ∈ {authoring-auto, data-derivable-auto, compute-bound-defer,
  external-defer, judgment-defer} and, for -auto findings, a structured fix_action
  (op / target / source_for_data / spec). This makes downstream remediation
  deterministic and fabrication-safe.
• Step 8 — consent-gated fix: after the report is written, if ≥1 auto-fixable finding
  exists, the reviewer asks the user (via AskUserQuestion) whether to fix now. It never
  modifies the ARA without an explicit first-person "yes", and -defer findings always go to
  the human worklist. Zero auto-fixable findings → no prompt.
• references/remediation.md — the Seal Level 3 contract the reviewer obeys strictly
  once authorized: the fabrication firewall (a fix may only re-express/relocate/scope/
  re-file existing knowledge — never invent a new empirical number; data-derivable-auto
  tables require a source_for_data), re-run Seal Level 1 after each round, anti-gaming (no
  score bump from self-remediation unless the epistemic ceiling actually moved), provenance,
  append-only journey, max 3 rounds.
• references/seal1_check.py — portable Seal Level 1 structural validator to run after
  each remediation round.

Why internalized (not a separate skill)

Per design discussion: folding remediation into the reviewer's references keeps the reviewer
self-contained and makes the "fix strictly per the contract" constraint unambiguous, instead
of depending on a separate installed skill.

Validation

• seal1_check.py runs clean (PASS 171 / FAIL 0) against a compiled ARA.
• Step 8 + reference wiring verified; no dangling standalone-skill references.

🤖 Generated with Claude Code

[AmberLJC]

Let's not go with a validator like seal1_check.py here — its checks are whole-file substring matches (not per-entity) and can't verify that data-derivable-auto numbers came from their cited source, so it gives a false sense of safety rather than real enforcement; can we deprioritize this skill for now?