data: add anti-saturation benchmark schema and readiness diagnostics (#419)

[AbdelStark]

Summary

Adds the RFC-0016 anti_saturation_semantic_v1 profile and the codelewm.downstream_anti_saturation_report.v1 diagnostic so a hard downstream reranking pack exposes anti-saturation eligibility before any CodeLeWM scoring. The change is additive: a new schema-only downstream_anti_saturation module plus optional profile / hard_negative_class config fields on the existing pack builder. Plain (non-profile) fixture configs are byte-for-byte unaffected.

Linked Issue

Closes #419.

Spec / RFC Reference

• Spec section: docs/spec/11-llm-world-model-harness.md
• RFC: docs/rfcs/RFC-0016-hard-downstream-reranking-benchmark.md

Public Surface Impact

New Python API (exported from codelewm.eval): build_anti_saturation_report, build_anti_saturation_claim_gate, compute_model_independent_baselines, validate_hard_negative_class, validate_anti_saturation_report, validate_anti_saturation_claim_gate, anti_saturation_report_json_schema, lexical_similarity, stable_random_key, DownstreamAntiSaturationError, and the ANTI_SATURATION_* / HARD_NEGATIVE_CLASSES / ceiling constants.

New schema versions (additive): codelewm.downstream_anti_saturation_report.v1, codelewm.downstream_anti_saturation_claim_gate.v1.

New config keys (both optional): profile (pack config, only anti_saturation_semantic_v1), hard_negative_class (candidate). DownstreamBenchmarkPackResult gains optional anti_saturation_report_path / anti_saturation_eligible. No existing field, baseline tuple, or schema version changed.

No new CLI command — the existing eval downstream-pack emits the report when the config sets the profile.

Validation

uv run pytest tests/eval/test_hard_downstream_schema.py tests/eval/test_hard_downstream_pack.py -q   # 21 passed
uv run pytest tests/eval/ tests/test_imports.py tests/harness/test_output_schemas.py -q              # 199 passed
uv run python -m compileall -q codelewm/eval/downstream_anti_saturation.py codelewm/eval/downstream_pack.py codelewm/eval/__init__.py
git diff --check

Artifact Impact

A pack built with profile: anti_saturation_semantic_v1 now writes reports/anti_saturation_report.json (schema codelewm.downstream_anti_saturation_report.v1), records it in manifest.json metadata (profile, anti_saturation_report, anti_saturation_eligible), and adds anti_saturation_eligible / anti_saturation_blocked_reasons to the readiness report. Packs without a profile are unchanged.

Deprecations

none

Caveats / Follow-ups

• The pack-time report computes the model-independent baselines (no-action/lexical/LLM-order/random). The CodeLeWM-scored baselines and the lift CIs are added at eval time in eval: score hard benchmark baselines and CodeLeWM claim gate #422.
• The bundled fixture has one task, so it is intentionally eligible=false (blocked only by problem_count_below_minimum:1<100), preserving the v1.0 claim boundary.
• Follow-ups: data: build public-safe hard-negative candidate pools #420 (hard-negative pool builder), harness: ingest LLM candidate packs into the hard benchmark #421 (LLM candidate ingestion), eval: score hard benchmark baselines and CodeLeWM claim gate #422 (scoring + claim gate), results: publish hard benchmark artifacts and claim audit #423 (publication).