[diegosouzapw]

Summary

• Preserve short events selected by experimental segment-aware sampling through the final caption budget.
• Let the structural sampler allocate the actual caption budget instead of doubling it and uniformly thinning away selected short events later.
• Bind result-cache identity to the corrected candidate-frame budget; reject results from the previous oversized policy.
• Keep uniform/scene-aware behavior, conservative dedup thresholds, frame caps, defaults and promotion allowlists unchanged.

Reproduction

The real FFmpeg rapid-cuts fixture contains black/white/black/white half-second events followed by a moving test pattern. With an eight-caption budget, structural sampling found all four short events, but the later 10-to-8 cap discarded the second black interval. The new public-helper integration test fails before the fix and passes afterward. A separate public-guardrail cache regression proves old metadata is not reused.

Validation

Executed on the dedicated VPS, not the development host:

• Actual-FFmpeg RED: missing [1, 1.5) short event.
• Cache identity RED: 16 !== 8.
• Focused affected suites: 84 passed, zero failures/skips.
• Post-commit broader exact-head video suites: 362 passed, zero failures/skips.
• Changed-file ESLint: PASS.
• npm run typecheck:core: PASS.
• Base-relative file/test size gate: PASS.
• Normal commit hooks passed; no hook bypass or baseline relaxation.

Head: 7ccf49155ab39a9f93beea33ee4f12e4dc32954f.
Base snapshot: 47688ba5ea28223cf240d4a3a1b542e249ad63b2.
The 18-arm frozen-corpus resource rerun retained every declared fact interval for all nine segment-aware cases, within the original CPU/RSS/pre-analysis caps. Its aggregate remains FAIL because the unchanged uniform baseline misses short events; sampled temporal coverage is not semantic/model-quality proof. Tests used existing VPS dependencies, so these receipts do not certify a clean-lock release build, packaging, deployment, corpus/model promotion, or current-base merge admission. Keep this PR draft until required exact-candidate gates are complete.

[github-actions]

CI Coverage Report

• Coverage job: skipped
• PR test policy: success

Coverage artifact was not available for this run.