fix(loudness): cue-out only from silence that extends to EOF — interior pauses are content (gh-#424)

[genwave-radio]

Fixes #424 — full analysis lives there. 🎙️✂️

🔍 What was wrong

FfmpegCueAnalyzer.ParseCuePoints took the last silence region as cue-out unconditionally. An interior pause that merely happened to be last — a 0.6s injected TTS sentence pause (they clear the 0.5s silencedetect floor), a quiet mid-track break — became liq_cue_out, and Liquidsoap truncated there. Net effect on air: the final sentence of every multi-sentence patter clip was cut, always at a pause, while booth logs, the cached file, and the preview all looked complete.

🔧 The fix (analyzer-level, not a TTS gate)

"Trailing" now means the region actually extends to EOF: open-ended (silence_start with no silence_end), or closed within 0.25s of the container duration. Interior silence never sets cue-out.

• Interior-pause-only files return CuePoints(cueIn, fileDuration) — not null — so DurationMs (F66.1) stays a real measurement for straddle boundary-fit and cross-duration planning.
• Entirely-silent files still return null (new head-guard covers the open-ended shape; spec'd).
• Chose the analyzer over a skip-cue-for-TTS gate in TtsSegmentSource because the defect is general — a music track with a quiet interior break mis-trims identically — and this keeps TTS leading-silence trim + measured durations. Misclassification degrades in the safe direction only: an untrimmed tail falls to the engine's blank.eat backstop, exactly the posture the class doc already names.

✅ Verification

• 3 new integration specs against real ffmpeg (interior pause → full extent; interior + trailing → trims only the tail; silence-only → null) + 2 TestMedia generators.
• Red-checked: the interior-pause spec fails on the pre-fix parser (Assert.InRange — cue-out at the pause start) and passes with the fix.
• dotnet build 0 warnings; full suite green: 3,602 passed / 0 failed across all 8 test projects.
• On-box confirmation after deploy: the playout: TTS patter airs truncated at the last pause — cue-out from silencedetect cuts the final sentence of every multi-sentence blurb #424 body's one-liner (silencedetect on a cached blurb) plus ears — back-announce/weather/lead-in should play to their final words.

🚚 Rollout notes

• TTS blurb/evergreen cues are re-measured per process (in-memory memo + file re-analysis), so playout self-heals on deploy — no cache sweep needed.
• Persisted cue rows keep their old values until re-measured: music catalog rows (re-enrichment) and authored imaging rows (SafeSegmentAuthor persists cue at authoring). Multi-sentence authored imaging created before this fix may still carry a truncating cue_out — re-author or shout if we want a follow-up migration.
• playout: TTS patter airs truncated at the last pause — cue-out from silencedetect cuts the final sentence of every multi-sentence blurb #424's riders deliberately not taken here (cache key omits SentencePauseSeconds; blank.eat 1.5s headroom vs the pause-setting range) — still open on the issue.