docs: lock hard downstream benchmark spec and tracker (#418)

[AbdelStark]

Summary

Lands the repo-level contract for the hard anti-saturation downstream reranking benchmark (RFC-0016), so future agents cannot weaken the anti-saturation gate. This is documentation + docs-tests only; no runtime/code surface changes. It is the first child of tracker #417 and sets the contract that #419–#423 implement.

Linked Issue

Closes #418.

Spec / RFC Reference

• Spec section: docs/spec/11-llm-world-model-harness.md (anti_saturation_semantic_v1 profile + codelewm.downstream_anti_saturation_report.v1)
• RFC: docs/rfcs/RFC-0016-hard-downstream-reranking-benchmark.md

Public Surface Impact

none (docs + docs tests only). No CLI flags, JSON schemas, manifest fields, error types, config keys, or Python APIs are added/changed/removed. The RFC names future schema strings (codelewm.downstream_anti_saturation_report.v1, anti_saturation_semantic_v1) but no code implements them yet.

Validation

uv run pytest tests/docs/
# 144 passed, 1024 subtests passed
uv run pytest tests/docs/test_hard_downstream_benchmark.py tests/docs/test_implementation_tracker.py tests/docs/test_final_paper_package.py -q

Artifact Impact

none on-disk at runtime. Documents the planned published artifact set and the codelewm.downstream_anti_saturation_report.v1 schema for later issues.

Deprecations

none

Caveats / Follow-ups

• This tracker does not reopen the broad coding-improvement claim. Public wording stays diagnostic until the locked benchmark gate (eval: score hard benchmark baselines and CodeLeWM claim gate #422/results: publish hard benchmark artifacts and claim audit #423) opens.
• Follow-ups: data: add anti-saturation benchmark schema and readiness diagnostics #419 (schema + readiness diagnostics), data: build public-safe hard-negative candidate pools #420 (hard-negative pools), 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 + claim audit).