fix: CI green again — Swift 6.2 type-checker limits, duplicated subtitle cues, AV1 test on VMs

[bilipp]

Summary

CI on main has been red since 069a818, and #39/#40 added more failures. None of it was a logic bug in those PRs. Three separate problems:

1. fix(ci): keep the sources within Swift 6.2's type checker

The workflow asks for Xcode_26.app, which the macos-15 image no longer has, so the fallback selects Xcode 26.3 (Swift 6.2). Local development uses Xcode 27 (Swift 6.4), and code that 6.4 accepts instantly, 6.2 rejects:

Where	6.2 error	Fix
Tests/.../Support/ChannelDrain.swift:88 (on main since 069a818)	initializer for conditional binding must have Optional type: inside the initializer of the local result, a bare result binds to that local	if let stored = self.result
DolbyVisionMapping.swift:174 Matrix3 (#39)	unable to type-check this expression in reasonable time: nested reduce(0) closures over untyped literals	plain loops, explicit types
DolbyVisionTests.extraction (#39)	2.4 s to type-check even on 6.4: #expect(pieces == [.polynomial(…)])	typed expected values
HEVCBaseLayerFilterTests hvcC builder (#40)	~0.1 s on 6.4	loop

I found the first slow spots by compiling with -warn-long-expression-type-checking=25 on 6.4. But 6.4's timings turned out to be a poor guide: CI's first run on this PR rejected three more test expressions that 6.4 checks in milliseconds. They're fixed by 84f8466:

• DolbyVisionTests' Reference.multiply/apply, the same reduce(0) shape;
• the HEVCBaseLayerFilterTests + chains over untyped literals.

Swift 6.2 isn't installed locally, so CI is the real check.

3. fix(test): don't take VideoToolbox's word on AV1 hardware

Once everything compiled, the AV1 test failed on CI: 72 of ~96 frames, all from software, although VTIsHardwareDecodeSupported(AV1) returned true. The runner is a virtualized Mac whose VideoToolbox claims AV1 and then fails it. The engine did what its recovery policy says:

• rebuild on dav1d,
• emit .downgradedToSoftware,
• resume at the next keyframe, one 24-frame GOP later.

The test now distinguishes three cases:

• no AV1 hardware claimed (the Apple TV case): dav1d from the first frame, nothing lost;
• hardware claimed and working: hardware frames;
• hardware claimed but failing: a reported downgrade that loses at most one GOP.

2. fix(subtitle): store a re-read cue once, not again

"embedded SRT track produces timed cues" failed intermittently, including on clean main: one run in 3 to 9 under full-suite load. Printing the store on a failing run showed every cue stored twice. That's an engine bug, not a test problem.

SubtitleStore is a timeline that deliberately survives seeks, but the decoder inserted every cue it decoded. Whenever the demuxer re-read subtitle packets it had already delivered, the cues were stored again:

• on every backward seek: users saw that range's lines doubled;
• in the backfill seek of a track selected late (selectSubtitleTrack), the test's path: when live delivery won the race against that seek, both copies landed.

insert() now drops a cue identical to one it already holds (same start, end and trimmed text). Different cues that share a start both stay.

Type of change

• fix — bug fix
• feat — new capability
• perf — performance improvement
• refactor — neither fixes a bug nor adds a feature
• docs — documentation only
• test — adding or fixing tests
• build / chore — build pipeline, tooling, or maintenance

Invariants

None affected: the only data-plane change is the dedup inside SubtitleStore.insert, under its existing lock.

• No blocking FFmpeg call was added to an actor, and no stage bypasses Channel.
• No rebuild-in-place: a new URL is still a new session object.
• Raw container timestamps still stop at TimestampUnwrapper.
• A/V sync is still owned by AVSampleBufferRenderSynchronizer (no new heuristics or clocks).
• No force unwraps, as!, fatalError, or assertionFailure in engine paths; failures are typed.
• No new global mutable state; configuration still flows through PlayerConfiguration.
• No control flow derived from av_log strings.

Verification

• swift build is warning-free. Nothing new warns; Render/SystemRenderer.swift:63-64 still has its two existing main-actor isolation warnings.
• swift test is green, six runs out of six (Swift 6.4, local dav1d xcframework):

Test run with 104 tests in 16 suites passed after 19.169 seconds.

• I added or updated tests for this change.
  • storeDeduplicates: a store unit test.
  • seekBackKeepsCuesUnique: reproduces the doubling deterministically.
  • Both fail with the dedup reverted (5 issues).
  • The AV1 decode test now separates "VT claims AV1" from "VT decodes AV1".
• New media fixtures: none.

Platforms verified

• macOS
• iOS / iPadOS: CI build job
• tvOS: CI build job
• visionOS

Not in this PR

• The workflow still asks for Xcode_26.app and silently falls back to the newest Xcode on the image, so the CI toolchain can change without notice. It may be worth pinning it explicitly, e.g. to Xcode_26.3.app, or to whatever minimum toolchain the package should support.