feat(curator): generalise integrity-repair escalation to all invariant kinds

[nickroci]

Summary

Generalises the just-merged broken-wikilink repair escalation so ALL three check_invariants kinds flow through the SAME enqueue -> Librarian-proposes -> Scholar-executes pipeline and the SAME in-flight concurrency guard, kind-dispatched (one mechanism, not three copy-pasted paths).

• overcap_dir — a flat dir over MAX_FLAT_DIR_ENTRIES enqueues a repair task; the Librarian proposes a split_folder / move_entry rebalance.
• bad_frontmatter — a missing / unparseable / incomplete frontmatter block enqueues a task; the Librarian proposes an update_entry that re-serialises valid frontmatter.
• Scholar-prompt audit — repair-originated proposals (packets carrying repair_fingerprints) are now explicitly verify-and-execute, not judged for novelty/dedupe, so the APPROVE-OR-VETO + salience-novelty framing can no longer wrongly veto a fix as "not novel"/"duplicate".

How it works

• scholar_prompt.check_invariants_detailed() returns structured InvariantViolations carrying (repair_kind, file, target, context); check_invariants() is now a thin string-list projection over it (public contract preserved).
• Over-cap and bad-frontmatter violations carry the fields to build a repair_queue.RepairTask. The Scholar's post-write _check_invariants_safe escalates each via _escalate_invariant_violation, reusing RepairQueue.enqueue + the in-flight fingerprint guard. Wikilinks (display-only here) still escalate via the dedicated deterministic repair pass; READMEs stay display-only (reconciler fixes them).
• The violation is left on disk, so it re-detects and re-escalates until fixed — no max-attempts, no give-up. The incoming packet's marker clears in review()'s finally, exactly as for wikilinks.
• The Librarian prompt's INTEGRITY-REPAIR section now dispatches on kind with per-kind repair recipes.

Test plan

• uv run --frozen ruff check . && uv run --frozen ruff format --check .
• uv run --frozen pyright (0 errors)
• uv run --frozen pytest — 583 passed, 90.45% coverage (>=90%)
• duplicate-code pylint gate — no findings (10.00/10)
• New tests: over-cap producer enqueues + in-flight guard; bad-frontmatter producer enqueues + guard; marker-release-then-reescalate; Librarian prompt handles the new kinds; Scholar-prompt repair-exemption framing; structured check_invariants_detailed projection + escalatable/display-only kinds.

🤖 Generated with Claude Code