feat(grading): add dormant older-problem credit support

[kamui]

Why

Add a separate “Older problems caught” count for true faults outside a change. A catch earns credit without creating a miss or changing ordinary recall.

This implements the recommended design as dormant support. The owner still needs to adopt the complete rules, reconcile affected references, approve case admissions and authorize grading. The five source faults remain candidates.

Part of #135. Refs #30.

Scope

• Add the older catalog, owner decisions and v3 verdict validation/mapping in bench/schema and bench/tools.
• Preserve native v2 outcomes and multiple known-problem matches in export v5.
• Add scorecard().older, the explorer panel, review details and CLI counts.
• Preserve pinned ruling extracts and add complete proposed rules, grading instructions and activation documentation.

Blast Radius

Touches current grading, export and scorecard consumers. Historical verdict readers remain supported, and tests keep bonus counts separate from ordinary detection, control failures and remedy safety. No active policy, saved judgments or raw benchmark evidence changed.

Verification

• Focused Python grading, adjudication, audit and export suites passed; v3 preparation/mapping exercised without paid dispatch.
• 68 web tests and TypeScript checks passed. Browser fixture inspection covered catches, zero catches and unavailable results.
• Real export and scorecard passed: 16 tasks, 14 configurations, 663 attempts; assessment coverage remains 0/623.
• Current grading, calibration, canonical claims and imported-evidence checks passed.
• Local review: Approved with complete coverage and no findings. Four exact source ruling copies retain trailing blank lines reported by the committed-range whitespace check.

[kamui]

Changes Requested (advisory) — 1 must-fix finding.

Findings

• [P1] [must-fix] Package the new validator dependency in both client probes — anchor bench/tools/grading_validation.py:10; fix bench/tools/grading_client_probe.py:45

Reviewed 1413c75. 53 changed files reviewed; four archived alternative plans and rationales ignored because they are superseded research evidence. Traced grading, export, scorecard and client probe consumers. Required independent verification completed.

Review details

Intent: Add dormant older-problem credit with separate counts, v3 grading and historical verdict export support.

Issue fit: The scoped support implements the proposed plan for #135. Policy adoption, reference reconciliation, case admissions and grading remain future activation work. Both client compatibility probes need their new validator dependency.

Reviewed: 1413c75 against merge-base b25a00a.

Checks:

• At 1413c75, the focused Python run covered 116 tests. Grading, export and audit checks passed; nine ruling checks lacked pinned ignored research responses. After restoring hash-verified inputs, all 42 ruling tests passed.
• At 1413c75, bun test ./src/lib/scoring.test.ts passed 37 tests. Live bun run dev:fixture inspection confirmed catches, complete zero catches, unresolved credit and missing assessment states.
• Both isolated probe module lists reproduce an import failure at head and succeed at base; adding older_grading.py to the disposable copy restores validation. Client compatibility and grading regressions CI both fail at their validate step.
• Exact-head Web and evidence CI passed 68 web tests, 27 export tests and TypeScript checking; its readable output was accepted.
• Independent safety checks held for owner admission, ordinary scores and historical export compatibility. The follow-up confirmed the dependency defect.
• The PR body reports other local checks without readable run artifacts; these were not reused. Publication uses kamui because no reviewing app is declared; gating remains withheld.

[kamui]

Addressed at fdb6fac19b1bb065c12d546c28b9f5f236a9ff75. 1 implemented in commit fdb6fac1.

• implemented: client-probe validator dependency. Both isolated bundles include the new dependency.

Checks on the clean committed tree at fdb6fac19b1bb065c12d546c28b9f5f236a9ff75:

• python3 -m unittest discover -s bench/tools -p test_grading_client.py: passed 1 test.
• python3 -m unittest discover -s bench/tools -p test_codex_grading.py, with Codex 0.160.0 as pinned in CI: passed 4 tests. The initial default Codex 0.162.1 run passed the actual-client probe but failed the unchanged pinned-catalog assertion.
• Both client probes exercised validation against local fake APIs with zero paid calls.
• bun run verify:import: passed, 2,775 source files and 288 archives verified, 9 superseded source-hash mismatches and 0 unavailable.
• git diff 1413c754..HEAD --check: passed for this resolution.

Independent review reproduced both original bundle failures and both fixed validation successes, found no introduced blocker, and checked the reply claims. Its scope was this two-line resolution; actual-client results above came from the recorded test runs. Unchanged web, export and scoring suites were not rerun in this round.

The thread is resolved. No questions or disputes remain from the inventoried feedback. This addressing round is finished and ready for re-review.

Re-requesting review from @kamui.

[kamui]

Approved (advisory) — no findings. Delta review of 1413c75..fdb6fac.

Reviewed fdb6fac. Reviewed both delta files and enclosing probe functions, carried the prior complete review for unchanged paths, and completed isolated verification of the prior must-fix and the isolation premise.

Review details

Intent: Re-review the two-client validator dependency fix in 1413c75..fdb6fac.

Issue fit: The resolution fixes the prior blocker in the dormant support for #135. Policy adoption, case admission, reference reconciliation and grading authorization remain later activation work.

Reviewed: fdb6fac against merge-base b25a00a.

Checks:

• At fdb6fac, AST-derived disposable bundles for both probes validate the exact empty-review fixture with exit 0; both prior-head bundles reproduce the missing older_grading import.
• Independent verification reproduced both failures and both fixes; the isolation premise holds for this module-copy delta.
• Exact-head Client compatibility CI passed its one installed-client test with all six grading tools exercised and ambient markers absent. Its log was read and accepted.
• Web and evidence CI is successful; Grading regressions CI was still running at the status capture and is not claimed as passed.
• The addressing reply reports four Codex 0.160.0 tests, one installed-client test and verify:import. Those prose-only results were not reused; unchanged scoring, export and browser checks retain their prior review disposition.
• Publication uses kamui because no reviewing app or token is declared; this self-review uses COMMENT and withholds gating.

Settled findings

• grading-validation/client-probe-dependency — fixed: Both copy lists include older_grading.py at fdb6fac; independently reconstructed isolated bundles validate successfully. The prior-head lists reproduce ModuleNotFoundError.