feat(episodes): materialize opt-in candidate v2 oracle

[stacknil]

Summary

• add an explicit window-separated-core-contrast-v2 evaluator path while preserving the default candidate-v1 writer byte for byte
• materialize exact minimum-span threshold-core admission evidence without hiding rejected selections
• version the research oracle schema, add a committed v2 golden, and lock positive, negative, cross-version, and permutation-compatible behavior

Contract change

• Affected producer: scripts/evaluate_episode_candidate.py research artifact only
• Migration: additive and opt-in through --algorithm window-separated-core-contrast-v2
• Backfill: none; existing v1 artifacts are not rewritten
• Compatibility: default and explicit v1 output match both committed v1 goldens byte for byte; Detector::analyze(), CLI findings, and loglens.report.v3 are unchanged

Review decision record

• Design decision: retain candidate-v1 selection as the evidence layer and materialize candidate-v2 admission as a separate, auditable policy projection
• Main risk: consumers could misinterpret the fixed 2x research ratio as production calibration
• Compatibility impact: none for v1 consumers; v2 is a new format bound to a distinct algorithm ID/version and required admission shape
• Rollback path: remove the v2 algorithm branch, v2 schema definitions, and v2 golden; the unchanged v1 writer and artifacts require no migration

Validation

• python -m unittest tests.test_episode_candidate_compatibility - 2 passed
• python -m unittest discover -s tests -p 'test_episode*.py' - 43 passed
• python -m ruff check . - passed
• v1 continuous, v1 isolated, and v2 continuous CLI golden replay - passed
• cmake --build build --config Debug - passed
• ctest --test-dir build -C Debug --output-on-failure - 7/7 passed
• gitleaks git --log-opts="origin/main..HEAD" --redact --no-banner - no leaks
• final diff review - 9 files; no runtime, public report, or build-contract files changed

Merge gate

Draft because this is a report/schema behavior change and the diff exceeds 500 lines including the committed golden. Keep open for delayed review; after CI passes, perform another diff review before marking ready.

[stacknil]

Delayed diff review completed after 14.25 hours.

The review found one schema permissiveness issue: candidate-v2 is defined around a fixed 2x admission ratio, but the schema previously accepted any positive rational. Commit 519ab1a pins the contract to {numerator: 2, denominator: 1} and adds a negative compatibility test.

Post-fix evidence:

• independent v1 byte compatibility: 2/2 passed
• episode test suite: 43/43 passed
• CTest: 7/7 passed
• Ruff and git diff --check: passed
• gitleaks: no leaks across 5 commits
• GitHub CI, Windows/Linux, CodeQL, Repo Sentinel, and fuzz smoke: all passed
• final diff review: no remaining findings; no src/, include/, CMake, CLI finding, or loglens.report.v3 changes

Design decision: candidate-v1 selection remains visible evidence; opt-in candidate-v2 admission is a separate policy projection.
Main risk: the fixed 2x research ratio could be mistaken for production calibration.
Compatibility impact: default and explicit v1 remain byte-for-byte stable; v2 is additive and strictly version-bound.
Rollback path: remove the v2 evaluator branch, v2 schema definitions, and v2 golden; v1 requires no migration.