for await never calls return() on an iterator whose next() threw, so the
error branches in mediaSamplesInRange / mediaSamplesAtTimestamps left the
iteration un-terminated: the pump's decoder callback kept queueing decoded
samples that no consumer or cleanup path could ever close, orphaning them
until GC ("A VideoSample was garbage collected without first being closed").
Set terminated (and ended, where the pump gates on it) before throwing in
both the InputDisposedError and out-of-band error branches of both sample
generators - samples emitted after the throw are then closed on arrival by
the existing checks in the decoder callbacks.
Observed on iOS Safari when the hardware VideoDecoder dies mid-iteration
("Decoding task did not complete"): 3-14 samples orphaned per failing
iterator; with this change, at most the single sample already in flight
during the error microtask window.
* Add non-owning conversions via ConversionOptions.ownsOutput
A conversion with ownsOutput: false only adds tracks to the output and
drives their media data; starting, finalizing, and metadata tags remain
the caller's responsibility. This lets multiple conversions and
directly-added user tracks compose on a single Output (see upstream
issue #436).
- ownsOutput: false allows a pre-populated output (state must still be
'pending') and seeds track-capacity accounting from existing tracks
- execute() requires the output to be started and never finalizes it
- cancel() closes only the conversion's own sources, releasing internal
synchronizer waiters, and leaves the output usable
- tags cannot be combined with ownsOutput: false
- isValid requires at least one contributed track instead of the
format's minimum track counts
Prototype for API discussion; default (owning) behavior is unchanged.
Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01AJdfnbY9AFh9i9dKgtrj6E
* Add external-audio example using a non-owning conversion
Demonstrates composing a user-owned audio track (synthesized voiceover
via OfflineAudioContext + AudioBufferSource) onto a picked video with
Conversion.init({ ownsOutput: false }), including progress reporting
and playback/download of the result.
Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01AJdfnbY9AFh9i9dKgtrj6E
* Release synchronizer waiters when canceling during output finalization
A non-owning conversion's cancel() previously no-oped entirely when the
output (owned by someone else) was already finalizing or finalized,
leaving pump loops parked in the track synchronizer and hanging
execute() forever. Now it still marks the conversion canceled and
releases parked waiters in that state, without force-closing sources
(finalization owns flushing them at that point).
Also adds coverage: non-owning onProgress monotonicity, canceling one
of two sibling conversions, capacity seeding across sequential inits,
exact metadata exclusivity, and cancel-before-execute.
Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01AJdfnbY9AFh9i9dKgtrj6E
* Document non-owning conversions on the converting-media-files guide page
Adds the doc section requested in #436: what ownsOutput: false does, the
required choreography (add tracks -> output.start() before execute() ->
run the conversion concurrently with your own sources -> finalize), the
cancellation split (conversion.cancel() leaves the output alive; cancel
both for a full abort and tear the output down on error paths), isValid
semantics in this mode, and the tags restriction with the
setMetadataTags() alternative. Also cross-links the fresh-output rule to
the new section. VitePress build passes with dead-link checking on.
Co-Authored-By: Claude Opus 4.8 <[email protected]>
* Clean up conversion logic, add Output.tracks and .hasEnoughTracks(), move new conversion tests around, remove external audio example
* non-owning -> composable, and update docs
* Update
---------
Co-authored-by: Claude <[email protected]>
Co-authored-by: Vanilagy <[email protected]>