[espetro]

Records the assessment of moving yt-dlp's bottlenecks into pyscriptc-compiled helpers that stock CPython yt-dlp imports, versus compiling the whole app. The note only; no code.

Verdict: not worth it on current evidence.

• 93.8% of a real offline yt-dlp run is not yt-dlp's own Python. The largest block (36 to 40% of a 1.4 to 1.7 s CLI run) is CPython's re compiling 1704 _VALID_URL patterns.
• The one real yt-dlp hotspot, orderedSet, is an O(n^2) dedup: 36.3 ms for 1754 items, 0.4 ms with a set. A one-line upstream fix beats a native rewrite.
• No importable-extension path exists: --emit binding makes a ctypes loader, not a PyInit_* module, and its ABI carries only int, float, bool and str. A hand-built PyInit_* prototype of parse_duration ran 1.4x to 2.4x slower than stock (a 1.49 us ABI floor per call plus a regex engine slower than CPython's C re).
• The native lane itself is fast on a typed orderedSet (24.8x), but that program cannot be called from CPython.
• Every candidate fences PS1001 because stock yt-dlp is unannotated, so this is an annotation wall, not a language wall.

Cut from the ytbottle session's branch (53 commits, much of it duplicating the ytnative chain) so the note lands cleanly from master.

🤖 Generated with Claude Code

[cloudflare-workers-and-pages]

Deploying pyscriptc with    Cloudflare Pages

Latest commit:	3da95d5
Status:	✅  Deploy successful!
Preview URL:	https://474b3beb.pyscriptc.pages.dev
Branch Preview URL:	https://work-ytbottle-note.pyscriptc.pages.dev

View logs