test(research): characterize uniform background volume

[stacknil]

Summary

• add a focused uniform-rate background control at the threshold/window boundary
• lock the transition from one selected candidate at 13 events to two at 14 events inside one activity segment
• document that candidate-v1 episode multiplication does not, by itself, prove multiple dense peaks

Why

The remaining qualitative deployment question was whether window-separated candidate multiplication was specific to separated density peaks. It is not: with threshold 5, an inclusive 600-second window, and one event every 150 seconds, 14 homogeneous events remain one v0.6 activity segment but candidate v1 selects line:1..line:5 and line:10..line:14.

This is the smallest decision-changing control at that cadence. The stopping rule was the first adjacent event count that changed the selected episode count (13 produces 1; 14 produces 2), so no broader parameter sweep was performed.

How to validate

• python -m unittest tests.test_episode_candidate_core.WindowSeparatedSelectionTests.test_uniform_threshold_rate_can_multiply_episodes_in_one_segment
• python -m unittest discover -s tests -p 'test_episode*.py'
• python -m ruff check scripts tests
• cmake --build build --config Debug
• ctest --test-dir build -C Debug --output-on-failure
• mutation check: disabling the multi-window predecessor link makes the focused control fail; restoring it passes
• gitleaks git --no-banner --redact --log-opts='origin/main..HEAD' .
• git diff --check origin/main...HEAD

Risk / rollout notes

• Design decision: treat this as a negative calibration result and block production adoption of candidate v1 without a density-contrast rule or explicit alert-volume budget.
• Main risk: overgeneralizing one synthetic cadence into a measured real-world false-positive or alert-volume rate.
• Compatibility impact: test and ADR only; Detector::analyze(), CLI behavior, report-v3, and released v0.6 behavior are unchanged.
• Rollback path: revert 41131d1 and 0e0568e.

[stacknil]

Post-CI final review complete.

• Checks: all 11 reported runs passed (Ubuntu, Windows, parser fuzz smoke, Repo Sentinel, and CodeQL).
• Design decision: the 13-to-14 event boundary is a negative calibration result; candidate v1 must not move into production unchanged without density contrast or an explicit alert-volume budget.
• Main risk: one sanitized uniform cadence must not be presented as a measured real-world false-positive or alert-volume rate.
• Compatibility impact: test and ADR only; no diff under src/, include/, CMakeLists.txt, or the research core implementation. Detector::analyze(), CLI, report-v3, and v0.6 behavior remain unchanged.
• Rollback path: revert 41131d1 and 0e0568e.
• Evidence: branch is 0 behind / 2 ahead, merge state is clean, 28/28 research tests and CTest 7/7 passed locally, the focused mutation was killed, and Gitleaks/privacy scans are clean.
• Stopping rule: the first adjacent event count at the fixed cadence that changes selected episode count was sufficient; no low-information parameter sweep was added.
• Reviews/comments: no bot or reviewer action items.