feat(engine,producer): apply VST chains in the offline audio mix

[vanceingalls]

Stack 3/5 — replaces #2602–#2605. Base: #2920.

A track carrying data-vst-chain is bounced through the VST host after it is trimmed and before volume automation is baked, so plugins see the dry signal and the envelope applies to their output.

A VST failure is deliberately fatal for the whole mix, not a per-track soft failure. Every other audio failure mode degrades gracefully (source missing, download failure, extract/prepare failure): the track drops, a failure is recorded, siblings continue. Silently substituting the untreated signal for a carved one ships a render that sounds plausible and is wrong — worse than a loud failure. So VstChainProcessingError escapes the per-element catch and rejects the whole call.

Because the per-element work races under Promise.all, that rejection does not wait for in-flight siblings. An internal AbortController (chained off the caller's signal) is threaded through extract, prepare and the bounce so a sibling's sidecar subprocess is aborted before workDir is removed from under it.

Lambda has no plugin host, so lambda render and lambda render-batch reject a composition with any VST chain up front, before any AWS-calling code runs.

Notes for review

• The 297 deletions are real edits to audioMixer.ts / audioMixer.test.ts against current main, not stack-internal churn. main grew a structured AudioProcessingFailure taxonomy after the original stack forked; the VST path was grafted onto that rather than onto the old string-error shape.
• render-batch was not guarded in the original feat(vst): voiceover spectral-carve VST pipeline (V1) #2602 — only render was. Same trust boundary, one line, so it is covered here.
• The guard's test was rewritten: it asserted the old process.exit contract, and main has since migrated the CLI to a throwing CliRuntimeError.

🤖 Generated with Claude Code

[vanceingalls]

Warning

This pull request is not mergeable via GitHub because a downstack PR is open. Once all requirements are satisfied, merge this PR as a stack on Graphite.
Learn more

• feat(studio): VST FX panel section behind STUDIO_VST_ENABLED #2923
• feat(studio,core,player): stream VST-processed audio in the studio player #2922
• feat(engine,producer): apply VST chains in the offline audio mix #2921 👈 (View in Graphite)
• feat(studio-server,cli): VST sidecar host and carve API #2920
• refactor(studio): share the flat-inspector test render harness across panel suites #2919
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.

[vanceingalls]

Superseded by the Web Audio implementation.

The sidecar approach is replaced by running Web Audio graphs on both ends: the studio previews in a live AudioContext, and the engine renders the same graphs in an OfflineAudioContext inside the headless browser it already drives. One implementation per effect, so preview matching render is a property of the architecture rather than a tolerance to police — which is what made the four effects that diverged under JUCE (compressor, limiter, chorus, phaser) stop being a problem.

That also retires the reasons this was cancelled in the first place: no spawned sidecar process, no external GPL repo, no uv tool install for users, and no plugin-missing failure mode in the render path.

Measurements behind the decision: https://claude.ai/code/artifact/219fd332-0542-4bcf-a4b7-7b7202e63fde

Branch left intact.