test(research): expose mean-gap padding sensitivity

[stacknil]

Summary

• add one focused characterization control for maximal-window padding sensitivity
• show that adding one in-window event preserves two selected dense cores but flips the 2x mean-gap diagnostic from pass to fail
• record why raw selected-window mean gap cannot be the sole candidate-v2 admission rule

Why

Hypothesis: adding evidence inside an already qualifying selected window should not make a two-dense-core admission diagnostic reject the pair when both dense cores and both selected episodes remain present.

The first adversarial control falsifies that hypothesis:

• base offsets: dense cores at 0..120 and 1260..1380 seconds, bridge events at 650 and 700
• base result: two five-event selections, 30-second internal means, 380-second bridge mean, 38/3x contrast, pass
• padded input: add one event at offset 600
• padded result: selections remain two episodes, the first expands to six events, internal maximum becomes 120 seconds, bridge mean becomes 220 seconds, 11/6x contrast, fail

The stopping rule is the first padding counterexample. No outlier-position or ratio sweep is needed.

Design decision

Do not use raw maximal-selected-window mean gap as a standalone candidate-v2 admission rule. Keep the current diagnostic as research characterization only and require a future design to separate threshold-sized dense-core evidence from maximal-window coverage.

This PR intentionally does not choose or implement a densest-core, quantile, trimmed-gap, or production policy.

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:

• candidate core: 10/10 passed
• episode research suite: 32/32 passed
• CTest: 7/7 passed
• build, Ruff, diff, privacy, and Gitleaks checks: clean
• fixture mutation: moving the padding event from 600 to 121 seconds made the focused test fail at the expected contrast assertion

Risk / rollout notes

• Main risk: one synthetic counterexample proves insufficiency but does not select a robust replacement statistic or estimate production alert volume.
• Compatibility impact: none; only a research test and ADR change. Detector::analyze(), CLI behavior, and loglens.report.v3 are unchanged.
• Rollback path: revert the two commits in this PR; no schema, data, or runtime migration is involved.
• Non-claim: the padded event is not classified as belonging or not belonging to a production episode.

[stacknil]

Post-CI final diff review completed.

• Design decision: reject raw maximal-selected-window mean gap as the sole candidate-v2 admission rule; do not choose a replacement statistic in this PR.
• Main risk: the synthetic counterexample proves insufficiency but does not estimate production alert volume or decide whether the padding event belongs to an episode.
• Compatibility impact: none; this PR changes only one research test and the ADR. Detector::analyze(), CLI behavior, and loglens.report.v3 remain unchanged.
• Rollback path: revert the two semantic commits; no schema, data, or runtime migration is involved.
• Evidence: 11/11 remote checks passed; local core 10/10, research suite 32/32, CTest 7/7, build/Ruff/diff/Gitleaks/privacy clean. Moving the padding event from 600 to 121 seconds made the focused test fail at its expected contrast assertion.
• Final review: selected ranges and the exact 38/3x to 11/6x derivation agree; no blocking comments, privacy issues, or unrelated runtime changes.