fix(media-use): reconcile disk-only narration WAVs into the audio ledger

[miga-heygen]

Summary

• A narration WAV can land in assets/voice/<id>.wav without ever going through this engine's own TTS synthesis loop: hand-placed, copied over from another project, or left behind by a run that wrote the file but didn't finish registering it. audio_meta.json's voices[] only ever grows inside that synthesis loop, so a file like this was invisible to every consumer that trusts voices[] (each workflow's assemble-index.mjs <audio> emission, captions.mjs) — the render succeeds and simply plays without that narration line, with zero warning.
• skills/media-use/audio/scripts/audio.mjs is the one shared engine backing this (its own header: "ONE implementation... workflows do NOT vendor a copy"), used by product-launch-video, faceless-explainer, and pr-to-video. Added a reconciliation step there — the single correct fix location — that scans assets/voice/ on every invocation (independent of --only, so a bgm/sfx-only re-run still catches drift) and backfills any file whose id isn't already known but is still requested by the current script with non-empty text. A stale file from a since-edited or since-cleared line is left alone, not resurrected.
• Placed before the existing hasVoice/totalDuration computation so a reconciled voice correctly participates in the BGM sizing/retrieval decision downstream, not just the final ledger output.
• Extracted the "does this line have real text" check (previously only inline in the synthesis loop) into a shared lineText() helper so the two call sites can't drift on that decision.

Test plan

• New test file with 5 cases, spawning the real engine script end-to-end (no mocks): a disk-only WAV gets backfilled with a real ffprobe-read duration; an already-registered voice isn't duplicated; a WAV with no matching current script line is left alone; a WAV for a line whose text was since cleared is left alone; a project with no assets/voice/ directory is an unaffected no-op.
• Confirmed genuine RED on the unmodified code (2 of the 5 cases fail with 0 backfilled voices instead of 1), then GREEN after restoring the fix.
• All 5 tests pass locally where ffmpeg/ffprobe are on PATH. In the Test: skills CI job specifically (no ffmpeg on that runner), these 5 tests currently {skip: !HAS_FFMPEG}, the same as the pre-existing BGM loop-fade test in this same skill — a pre-existing gap in that job's own design, not something introduced by this change.
• All 3 consuming workflows' own test suites (194 tests) and the full media-use test tree (351 tests) pass with no regressions.