fix(engine): recover audio duration via full packet scan when the container header has none

[miga-heygen]

Summary

extractAudioMetadata() previously trusted only ffprobe's container-header duration (format.duration) when probing an audio file. A container whose encoder never got a chance to seek back and patch that field in — e.g. audio piped to a non-seekable destination, which is exactly what a browser MediaRecorder-style capture does — comes back with no duration at format OR stream level, not merely zero. With no fallback, the render pipeline treated the track as zero-length and silently dropped it from the composition: no error, no warning, just missing audio in the output.

The function already had one precedent for recovering an untrustworthy duration — an AAC-LC-specific packet-count recomputation — but it's gated to that one codec/profile combination. This generalizes the idea to any codec: when duration is still unrecoverable after the existing logic, prefer the already-probed stream-level duration if it's usable, otherwise scan every packet's timestamp to EOF and take the last packet's pts_time + duration_time.

Only the true last parsed line is trusted (not a max across every line), mirroring the existing final-video-frame-timestamp prober already in the same file, which uses the identical output-retention settings for the identical reason: the retention limit that bounds a long probe's cost keeps only the tail of ffprobe's output, and that can slice through the middle of an early line — but never through the true last one. An earlier version of this fix took the max across all parsed lines instead, which a caught-in-review adversarial pass showed could let a truncated, numerically-bogus early line silently win over the correct answer.

Test plan

• New regression suite covering: successful recovery from a full packet scan, preferring the cheap stream-level duration when present (skips the scan entirely), graceful no-op (stays at 0, never throws) when the scan finds nothing or the probe itself fails, tolerating codecs/containers that omit per-packet duration_time, a truncated-leading-line case proving a bogus large timestamp is correctly ignored in favor of the true last line, and confirming the new fallback is skipped when the existing codec-specific refinement already recovered a duration.
• Verified the recovery mechanism against a real repro: piped a libopus-in-WebM encode to a non-seekable destination (ffmpeg ... -f webm - > out.webm) and confirmed via ffprobe that neither format- nor stream-level duration is present, while a full packet scan correctly recovers the true duration from packet timestamps.
• node --test/vitest run on the touched test file: 118/122 pass (4 pre-existing, unrelated failures traced to an unresolved Git LFS pointer for a fixture PNG in this environment — reproduced identically on a clean checkout of the base branch before this change).
• Downstream consumers (audioMixer, chunkEncoder, htmlCompiler, audioPadTrim test suites): all pass, no regressions.
• tsc --noEmit, oxlint, oxfmt, and fallow audit against the base branch: clean.