feat(research): add episode density-contrast gate

[stacknil]

Summary

• add a pure research diagnostic that requires every adjacent selected-window bridge mean gap to be at least 2x the larger internal mean gap
• use exact integer-microsecond Fraction arithmetic, binary-search bridge counts, and fail-closed candidate validation
• lock dense-peak, uniform-background, single-selection, invalid-ratio, and permutation controls; record the evidence boundary in the ADR

Why

Candidate v1 can split a uniform threshold-rate stream into two episodes, so episode multiplication alone does not prove two dense peaks.

Hypothesis: a 2x mean-gap contrast separates the existing continuous two-peak fixture from the known uniform-background false split.

Decision rule:

• accept this diagnostic as a candidate-v2 research prerequisite only if the dense fixture passes, the uniform control fails, and the single-selection case remains valid
• stop after those positive, negative, single, and invalid-input controls; do not sweep cadence values

Observed bounded evidence:

• continuous two peaks: 30 s internal means, 540 s bridge mean, 18x contrast
• uniform background: 150 s internal means, 150 s bridge mean, 1x contrast

Design decision

Keep candidate selection unchanged and evaluate its selected windows through a separate research-only diagnostic. Pair comparison is:

bridge_mean_gap >= 2 * max(left_internal_mean_gap, right_internal_mean_gap)

This is not wired into Detector::analyze(), the CLI, or loglens.report.v3.

How to validate

python -m unittest tests.test_episode_candidate_core
python -m unittest discover -s tests -p 'test_episode*.py'
python -m ruff check scripts tests
cmake --build build --config Debug --target ALL_BUILD -- /m:4 /nr:false /v:q /nologo
ctest --test-dir build -C Debug --output-on-failure
git diff --check origin/main...HEAD
gitleaks detect --source . --no-banner --redact --verbose

Local results:

• focused core: 9/9 passed
• episode research suite: 31/31 passed
• CTest: 7/7 passed
• Ruff and diff checks: clean
• Gitleaks: no leaks found
• mutation control: bypassing the contrast comparison made the uniform-background regression fail

Risk / rollout notes

• Main risk: 2x is an intentionally uncalibrated research threshold and could over- or under-separate unseen traffic shapes.
• Compatibility impact: none for stable v0.6 behavior; only research Python, tests, and the ADR change.
• Complexity: ordered validation and binary-search bridge counts add O(E log E + S log E) diagnostic work.
• Rollback path: revert the three commits in this PR; no migration or report-schema rollback is required.
• Non-claim: this does not estimate a real false-positive rate or authorize production adoption.

[stacknil]

Post-CI final diff review completed.

• Design decision: keep candidate-v1 selection unchanged and expose one pure research-only gap-contrast diagnostic.
• Main risk: the 2x ratio is not calibrated for production traffic and must not be treated as a production threshold.
• Compatibility impact: none; Detector::analyze(), CLI behavior, and loglens.report.v3 are unchanged.
• Rollback path: revert the three semantic commits; no data or schema migration is involved.
• Evidence: 11/11 remote checks passed; local core 9/9, research 31/31, CTest 7/7, Ruff/diff/Gitleaks clean; mutation bypass was caught by the uniform-background regression.
• Final review: no blocking review findings, privacy issues, or unrelated production-path changes.