fix(audio): abort YouTube downloads at the canonical 100 MiB ceiling

[cursor]

Summary

Continues #866 / #781. Native local-file and post-success YouTube bootstrap already used the 100 MiB policy, but download_youtube_audio still downloaded with no yt-dlp max_filesize and then rejected > 50MB. That rejected 60–100 MiB imports the new policy would accept and let a large transfer fill the cache root before the native check ran.

This branch keeps #866's policy stack and adds download-time admission:

• max_filesize and a progress hook abort in-flight once encoded bytes exceed DEFAULT_MAX_ENCODED_FILE_BYTES.
• Announced filesize / filesize_approx over the ceiling rejects before download=True.
• Post-download AudioResourcePolicy.validate_encoded_file_bytes deletes leftovers.
• User-facing oversize copy matches the local-file 100 MiB next-action message.
• Doctoring residual-risk text no longer claims Rust intake is missing; yt-dlp evidence is cited in APA 7th.

A 60 MiB YouTube import is now accepted; a file one byte over 100 MiB is not.

Verification

• Focused: PYTHONPATH=src python3 -m pytest tests/test_youtube.py tests/test_audio_resource_policy.py — 47 passed (26 YouTube + policy).
• ruff check / ruff format --check on the YouTube module and tests.
• mypy --strict --follow-imports=skip on youtube.py.
• ./scripts/harness/quickcheck.sh (full harness not run in this environment; exact-head CI must pass before Ready).

Security Notes

Attack surface

YouTube URL intake, yt-dlp subprocess/download, cache-root writes, announced filesize metadata, progress-hook byte counters, post-download file deletion.

Trust boundary

Renderer JSON is not authoritative. URL allowlist stays first. Encoded-byte admission is the canonical Python policy. Native bootstrap still re-checks filesystem-observed length before storing project state.

Mitigations

• max_filesize + progress-hook abort so unknown-size transfers cannot fill the cache root.
• Pre-download reject on announced oversize; post-download policy revalidation and delete.
• Payload-free errors (no paths, URLs, cookies, or audio content).
• Boolean/non-finite announced sizes are ignored rather than trusted.

Test points

• 60 MiB accepted (legacy 50 MB would reject).
• Exact 100 MiB accepted; 100 MiB + 1 deleted.
• Announced filesize / filesize_approx oversize never starts download.
• Progress hook oversize and yt-dlp max-filesize errors map to size_exceeded without leaking raw messages.

Dependency and Supply Chain

• No new direct dependency was added
• If a new dependency was added, this PR explains why it is needed
• runtime / dev / build / test classification is recorded
• alternatives were considered
• maintainer trust and update health were checked
• license fit was checked
• known security issues were checked
• transitive footprint impact was considered
• SBOM or supplemental inventory impact was recorded

i18n impact

• No user-visible string changed
• Korean and English locale impact was updated

The YouTube oversize engine message now matches the existing local-file 100 MiB copy already used by the desktop bridge. Locale files do not own this string.

Reviewer checklist

• Gitflow target branch is develop
• protected-branch rules were not weakened
• required checks are expected to stay green

Stacked on #866 / #865. Do not merge until those ancestors are in protected develop or this head is reconstructed onto the resulting base. Refs #781.

  

[seonghobae]

Superseded by the canonical existing #781 owner #866 after exact-head reconciliation. #866 now has merge head 1f3fdb8bfe869d747deaf4ce5cb53c23329e690f with #871 exact head 7b3b1e75ea8486869d0cc887122c9999ae2667fe as an explicit parent. compare(7b3b1e75…, 1f3fdb8…) is ahead by two commits and the only tree delta is #866's separate fail-closed coverage regression file services/analysis-engine/tests/test_audio_resource_policy_coverage_regressions.py; therefore every #871 source/test/doc/CHANGELOG change is preserved unchanged in #866. Closing this duplicate keeps one canonical writer for the audio-resource-policy boundary. Exact-head CI/review on #866 must still prove the integrated result before any Ready/merge classification.