fix(skills): loop-extended BGM writes .mp3, not MP3-in-a-.wav

[miga-heygen]

Summary

• ensureBgmCovers() — duplicated across skills/{faceless-explainer,pr-to-video,product-launch-video}/scripts/assemble-index.mjs — always re-encodes a short BGM track with libmp3lame when loop-extending it to cover the video's full duration, but named the output by preserving the source asset's own extension. A bgm.wav source therefore produced bgm.loop.wav containing genuine MP3 audio: a real container/codec mismatch.
• Confirmed via a live repro: ffprobe/ffmpeg/real Chromium (<audio> via CDP) all decode this file correctly today, because ffmpeg's WAV muxer tags it WAVE_FORMAT_MPEGLAYER3 and every consumer in HyperFrames' own pipeline goes through ffprobe/ffmpeg. So the originally-reported symptom (wrong/short duration, silent tail) does not reproduce with the current toolchain — this is a landmine for any future/adjacent naive WAV parser (e.g. Python's stdlib wave module, which hard-fails on format tag 0x55), not a currently-triggered bug.
• The function's own doc comment already promised the fix: "...into a sibling *.loop.mp3 and return that path" — the code just didn't match its own comment.
• Fix: always emit a .loop.mp3 path regardless of the source extension.
• music-to-video's assemble-index.mjs does not have this function — it mounts the raw BGM asset directly with no loop-extension step — so only 3 of the 4 skills needed the change.

Explicitly out of scope

• The Kokoro Japanese phonemizer garble reported alongside this — unverified without live model synthesis, not touched.
• The separate langApplied diagnostic gap — different piece of work.

Test plan

• Added a new regression test to all 3 affected assemble-index.test.mjs files ({ skip: !HAS_FFMPEG }, following the existing dither.test.mjs guard pattern): generates a real 1s WAV tone via ffmpeg -f lavfi, runs the actual assemble script against a storyboard whose TOTAL (3s) forces the loop-extend branch, then asserts the emitted <audio id="el-bgm"> src ends in .mp3 (not .wav) and that the file at that path really does decode as codec_name=mp3 via ffprobe.
• Confirmed the new test fails against the pre-fix code (reverted locally, re-ran, got AssertionError: bgm src should end in .mp3, got "assets/bgm/bed.loop.wav") and passes against the fix.
• node --test on all 3 modified test files: 12/12 pass.
• bunx oxlint + bunx oxfmt --write on all 6 changed files: clean, no reformatting needed beyond the authored diff.
• bunx fallow audit --base origin/main --fail-on-issues: clean, 0 issues across 6 changed files.
• Pre-commit hooks (lefthook: largefiles / tracked-artifacts / skills-manifest / commitlint) all passed.

🤖 Generated with Claude Code