Publish v0.9 claim audit

[AbdelStark]

Summary

• publishes the v0.9 final claim audit, public artifact index, dataset card, and seed-specific model cards
• adds the checked-in v0.9 eval artifact tree for seed 42 and seed 1729 across retrieval, surprise, probe, HumanEval/MBPP-Plus rerank, and p-pass calibration
• fixes rerank completion score rows to retain benchmark_id and split metadata so combined p-pass calibration does not collapse benchmark slices to unknown
• updates roadmap/goal prompt state so [TRACKER] v0.9 data/eval repair for cross-benchmark correctness claims #385/v0.9 hygiene: reconcile stale trackers and roadmap queue state #386-v0.9 eval/report: run full gate suite and publish claim audit #392 are historical completed diagnostic work, not an active queue

Claim Boundary

v0.9 stays overall claim-closed. Both seeds clear HumanEval WS-D reranking, but MBPP-Plus has zero lift over no-action, broad semantic-decoy and representation gates remain closed, and p-pass calibration is reported from completion-score baselines rather than a standalone serialized p_pass key.

Validation

• uv run pytest tests/docs/test_hf_ml_intern_training.py -> 8 passed
• uv run pytest tests/ -> 967 passed, 8 skipped, 1 torch nested-tensor warning
• uv run python -m compileall -q -x 'tests/fixtures/codestate/invalid_(before|after)\.py$' codelewm tests
• uv run codelewm --help
• uv run codelewm openrouter byok-register --dry-run --json
• uv run scripts/llm-world-model-demo -> manifest verify ok, artifact/html secret scans ok, claim gate closed
• parent-aware manifest verification for the v0.9 pack, both training runs, and all 12 checked-in v0.9 eval reports -> ok
• uv run codelewm secret-scan docs/benchmark/v0_9 --json -> ok, no findings
• git diff --check

Closes #392.