Merge upstream AetherEngine 7.36.0 into the Silo fork

.github/workflows/ci.yml

@@ -58,6 +58,15 @@ jobs:
         if: ${{ !cancelled() && steps.swift-tests.outcome != 'skipped' }}
         run: swift test --skip-build --filter "$AUTHORIZATION_TEST_SUITES"
 
+      # The same Swift Testing suite on a ONE-thread cooperative pool. Anything that blocks a pool
+      # thread (a sync read, a semaphore, a `Task.detached` around engine I/O) starves every other
+      # test here deterministically, instead of a different victim timing out on some later run of
+      # the 3-core runner. See docs/architecture.md, "Blocking work and the cooperative pool".
+      - name: swift test (one-thread cooperative pool)
+        env:
+          LIBDISPATCH_COOPERATIVE_POOL_STRICT: "1"
+        run: swift test --skip-build --disable-xctest
+
       - name: Software VOD packet-cache regressions
         run: |
           bash Scripts/test-software-packet-coverage.sh

Examples/README.md

@@ -18,7 +18,7 @@ Three samples covering different audiences:
    ```
    https://github.com/superuser404notfound/AetherEngine
    ```
-   Dependency Rule: Up to Next Major Version, starting from `7.28.3`. Add the `AetherEngine` library product to your app target.
+   Dependency Rule: Up to Next Major Version, starting from `7.36.0`. Add the `AetherEngine` library product to your app target.
 
 3. **Drop the file in.** Replace the Xcode template's default `App.swift` (or whatever the generated `@main` file is called) with the contents of [`MinimalPlayerApp.swift`](MinimalPlayer/MinimalPlayerApp.swift). The file is self-contained: it defines both the `@main App` struct and the `ContentView`.
 

Package.swift

@@ -38,7 +38,7 @@ let package = Package(
         // rebuilds still reach existing tags, which is where a pure rebuild
         // belongs; anything that adds slices or enables a component is a minor
         // and reaches consumers through an engine release.
-        .package(url: "https://github.com/superuser404notfound/FFmpegBuild", .upToNextMinor(from: "3.6.0")),  // 3.6.0: an MPEG-TS audio PID is identified by its payload when its label does not name the codec: the raw dts / truehd / loas demuxers are in, so the mpegts content probe can confirm DTS, TrueHD and LATM at all (without them the lenient mp3 probe named the track, whatever the label), and a PID labelled 0x03 (MPEG-1 audio) gets the same content probe as 0x04; a DTS-HD IPTV channel opened as mp3 and never produced a frame (#641); same n8.1.3 otherwise; 3.5.0: FFmpeg n8.1.3, 268 upstream commits over n8.1.2, of which six touch code this engine ships: VP9 kept `next_refs[]` referenced across `avcodec_flush_buffers()`, so a worker could reseed its references from the pre-flush set and the first inter frame after a seek decoded against released buffers (heap OOB read and write, and this engine decodes VP9 with frame threading and flushes on every reposition); the Dolby Vision RPU parser trims trailing zero padding again, which n8.1.2 had stopped doing, so a padded RPU no longer fails the ext-block size check and loses its dynamic metadata, and the RPU partition counts are bounded; a VC-1 picture ending on a one-bit skipped macroblock keeps that macroblock instead of falling into error concealment; the `dca_core` bsf clears the profile it has just stripped; and a container channel layout survives a failed decoder open during probing. Same dav1d 1.5.4, zimg 3.0.6 and libzvbi 0.2.45 (all current upstream), same component set, same six local patches, unchanged soname majors; 3.4.0: FFmpeg learns the `dav1` sample entry in both directions, the AV1 variant of `av01` that signals Dolby Vision, so a Profile 10.0 source can be read at all and can be remuxed with the tag its packaging requires; stock FFmpeg has it in neither of its two mp4 tag tables, so the read probes as "unknown codec" and the write fails EINVAL (#547); same n8.1.2 binaries as 3.3.0 otherwise, compiled against the Xcode 27 SDKs with unchanged deployment targets; 3.3.0: every slice a shipped app can embed carries its dSYM inside the xcframework, so Xcode copies it into `.xcarchive/dSYMs` and a crash inside FFmpeg symbolicates instead of staying a list of addresses (FFmpegBuild#4); same n8.1.2 binaries as 3.2.1 apart from their UUID, the libraries just compile with `-gline-tables-only` now and the debug map is harvested before the shipped binary is stripped; 3.2.1: vc1_parser seeds its parse context from extradata, so a seek landing on an entry point alone keeps its picture size instead of falling back to 0x0 or to a size read out of the following payload (#490, FFmpeg PR 24458); 3.2.0: the legacy Flash tail, video and audio together (flv1 / vp6 / vp6f / vp6a, nellymoser, adpcm_swf, speex, and FLV's pcm_s16be / pcm_u8 / pcm_alaw / pcm_mulaw), so a pre-2008 .flv plays with sound where before only H.264-in-FLV did; flashsv stays out, it needs zlib and would have been dropped silently; 3.1.0: native .wmv / .asf, the asf demuxer plus every WMA decoder (Standard, Pro, Lossless, Voice), which is the container half of the legacy Microsoft support 2.4.3 started (FFmpegBuild#3); 3.0.0: every target, product, framework bundle and install name carries an `Aether` prefix, so this build can sit in an app that already has an FFmpeg (FFmpegKit, MobileVLCKit, mpv); same n8.1.2 binaries as 2.5.0, only names changed; 2.5.0: libzvbi 0.2.45 for GHSA-86rm-g7qf-j2fh (OOB read/write + integer underflow, reachable through the libzvbi_teletext decoder), dav1d 1.5.4, zimg 3.0.6, and the concat demuxer removed (a script demuxer selectable by probing alone, which made any stream a potential file-open primitive); 2.4.3: the legacy Microsoft video decoders (msmpeg4v1/v2/v3, wmv1/wmv2/wmv3; FFmpegBuild#3); 2.4.2: pgssubdec missing-palette recovery, replaces the 2.1.1 Epoch-Continue retain (#142); 2.4.1: qtrle decoder; 2.4.0: visionOS (xros) device + simulator slices; 2.3.0: webvtt demuxer (standalone .vtt sidecars, plus the cue settings as packet side data); 2.2.0: matroska TTS warn-only per RFC 9559 (#145 rework); 2.1.3: sup demuxer (raw PGS sidecars); 2.1.2: matroska TrackTimestampScale clamp (#145, dropped in 2.2.0); 2.1.1: pgssubdec Epoch-Continue retain (#142); 2.1.0: yadif_videotoolbox + hwupload (Metal GPU deinterlace); 2.0.0: dynamic frameworks (LGPL), zvbi GPL excision
+        .package(url: "https://github.com/superuser404notfound/FFmpegBuild", .upToNextMinor(from: "3.7.1")),  // 3.7.1: the mp4 muxer no longer aborts the process when a fragment is flushed with a packet still queued in the interleaver behind the samples already written for its track (check_pkt had clamped them past it, so get_cluster_duration hit av_assert0(next_dts >= 0)); same n8.1.3 otherwise; 3.7.0: the mp4 muxer's dec3 box describes a DD+ track whose height channels and Atmos objects live in a dependent substream (Blu-ray style AC-3 core plus E-AC-3 dependent frame): chan_loc is derived from chanmap per TS 102 366 (upstream f10fdd6310, backported) and the TS 103 420 JOC extension is taken from the dependent substream too, so such a track is no longer described as a bare 5.1 bed and an Atmos receiver no longer reports multichannel PCM (#728); same n8.1.3 otherwise; 3.6.0: an MPEG-TS audio PID is identified by its payload when its label does not name the codec: the raw dts / truehd / loas demuxers are in, so the mpegts content probe can confirm DTS, TrueHD and LATM at all (without them the lenient mp3 probe named the track, whatever the label), and a PID labelled 0x03 (MPEG-1 audio) gets the same content probe as 0x04; a DTS-HD IPTV channel opened as mp3 and never produced a frame (#641); same n8.1.3 otherwise; 3.5.0: FFmpeg n8.1.3, 268 upstream commits over n8.1.2, of which six touch code this engine ships: VP9 kept `next_refs[]` referenced across `avcodec_flush_buffers()`, so a worker could reseed its references from the pre-flush set and the first inter frame after a seek decoded against released buffers (heap OOB read and write, and this engine decodes VP9 with frame threading and flushes on every reposition); the Dolby Vision RPU parser trims trailing zero padding again, which n8.1.2 had stopped doing, so a padded RPU no longer fails the ext-block size check and loses its dynamic metadata, and the RPU partition counts are bounded; a VC-1 picture ending on a one-bit skipped macroblock keeps that macroblock instead of falling into error concealment; the `dca_core` bsf clears the profile it has just stripped; and a container channel layout survives a failed decoder open during probing. Same dav1d 1.5.4, zimg 3.0.6 and libzvbi 0.2.45 (all current upstream), same component set, same six local patches, unchanged soname majors; 3.4.0: FFmpeg learns the `dav1` sample entry in both directions, the AV1 variant of `av01` that signals Dolby Vision, so a Profile 10.0 source can be read at all and can be remuxed with the tag its packaging requires; stock FFmpeg has it in neither of its two mp4 tag tables, so the read probes as "unknown codec" and the write fails EINVAL (#547); same n8.1.2 binaries as 3.3.0 otherwise, compiled against the Xcode 27 SDKs with unchanged deployment targets; 3.3.0: every slice a shipped app can embed carries its dSYM inside the xcframework, so Xcode copies it into `.xcarchive/dSYMs` and a crash inside FFmpeg symbolicates instead of staying a list of addresses (FFmpegBuild#4); same n8.1.2 binaries as 3.2.1 apart from their UUID, the libraries just compile with `-gline-tables-only` now and the debug map is harvested before the shipped binary is stripped; 3.2.1: vc1_parser seeds its parse context from extradata, so a seek landing on an entry point alone keeps its picture size instead of falling back to 0x0 or to a size read out of the following payload (#490, FFmpeg PR 24458); 3.2.0: the legacy Flash tail, video and audio together (flv1 / vp6 / vp6f / vp6a, nellymoser, adpcm_swf, speex, and FLV's pcm_s16be / pcm_u8 / pcm_alaw / pcm_mulaw), so a pre-2008 .flv plays with sound where before only H.264-in-FLV did; flashsv stays out, it needs zlib and would have been dropped silently; 3.1.0: native .wmv / .asf, the asf demuxer plus every WMA decoder (Standard, Pro, Lossless, Voice), which is the container half of the legacy Microsoft support 2.4.3 started (FFmpegBuild#3); 3.0.0: every target, product, framework bundle and install name carries an `Aether` prefix, so this build can sit in an app that already has an FFmpeg (FFmpegKit, MobileVLCKit, mpv); same n8.1.2 binaries as 2.5.0, only names changed; 2.5.0: libzvbi 0.2.45 for GHSA-86rm-g7qf-j2fh (OOB read/write + integer underflow, reachable through the libzvbi_teletext decoder), dav1d 1.5.4, zimg 3.0.6, and the concat demuxer removed (a script demuxer selectable by probing alone, which made any stream a potential file-open primitive); 2.4.3: the legacy Microsoft video decoders (msmpeg4v1/v2/v3, wmv1/wmv2/wmv3; FFmpegBuild#3); 2.4.2: pgssubdec missing-palette recovery, replaces the 2.1.1 Epoch-Continue retain (#142); 2.4.1: qtrle decoder; 2.4.0: visionOS (xros) device + simulator slices; 2.3.0: webvtt demuxer (standalone .vtt sidecars, plus the cue settings as packet side data); 2.2.0: matroska TTS warn-only per RFC 9559 (#145 rework); 2.1.3: sup demuxer (raw PGS sidecars); 2.1.2: matroska TrackTimestampScale clamp (#145, dropped in 2.2.0); 2.1.1: pgssubdec Epoch-Continue retain (#142); 2.1.0: yadif_videotoolbox + hwupload (Metal GPU deinterlace); 2.0.0: dynamic frameworks (LGPL), zvbi GPL excision
         // Pure-Swift SMB2 client (MIT) that speaks the protocol over
         // NWConnection. Replaces AMSMB2/libsmb2, which EPERMs on tvOS/iOS.
         // Pinned to the 0.3.x minor: SMBClient is pre-1.0 with an actively
@@ -53,9 +53,14 @@ let package = Package(
         // floor instead of backing off, so all of 5.x stopped resolving.
         .package(url: "https://github.com/superuser404notfound/LibDovi", .upToNextMinor(from: "2.1.0")),  // 2.1.0: dolby_vision 3.4.0, header additive only (two new CMv4.0 metadata entry points, nothing removed); 2.0.0: visionOS (xros) device + simulator slices, declared tvOS floor corrected to 17.0 (was published as 1.1.0, withdrawn: a floor raise is breaking and broke every 5.x pin that floated onto it); 1.0.2: iOS slices + x86_64 (Intel Macs)
         // TrueHD decoder that also decodes the Atmos object presentation (beds, objects, OAMD) for
-        // `objectAudioRendering`. The Rust `truehd` crate (Apache-2.0) behind a C API, prebuilt as an
-        // xcframework like LibDovi so consumers need no Rust. Pinned to the minor for the same reason.
-        .package(url: "https://github.com/superuser404notfound/LibTrueHD", .upToNextMinor(from: "2.0.0")),  // 2.0.0: SiloObjectAudio renamed after its transfer (AE#731): package LibTrueHD, product, module and xcframework TrueHD, same decoder and C API; 1.0.0: first release, as SiloObjectAudio
+        // `objectAudioRendering` (#731). The Rust `truehd` crate (Apache-2.0) behind a C API, prebuilt
+        // as an xcframework like LibDovi so consumers need no Rust. Pinned to the minor for the same reason.
+        .package(url: "https://github.com/superuser404notfound/LibTrueHD", .upToNextMinor(from: "2.3.0")),  // 2.3.0: decodes about a third faster again (harletty's truehd speed work: blocks decoded in place, table-driven Huffman, vectorised prediction filters, whole-block matrices; truehdd#42 included), byte-identical output; 2.2.1: rust_eh_personality exported weak, so it coalesces with LibDovi's copy and a host with its own Objective-C and C++ stays within compact unwind's three personality routines (#752); same decoder; 2.2.0: decodes about 30% faster (the truehd crate with its block and PCM buffers sized to the samples they hold, truehdd#42, from a fork of upstream main until that is merged), byte-identical output; 2.1.0: x86_64 in the iOS simulator slice, so a generic iOS Simulator build of a host links; same decoder; 2.0.0: transferred from Silo-Server/SiloObjectAudio and renamed (product TrueHD), rebuilt from source; same decoder as SiloObjectAudio 1.0.0
+        // DTS:X decoder for `objectAudioRendering`: lossless DTS-HD MA with the encoder's height and object
+        // folds removed from the bed, the fixed heights, and the alternate presentations' objects placed in
+        // the same room coordinates as LibTrueHD's. The Rust `dca` crate (Apache-2.0) behind a C API,
+        // prebuilt as an xcframework like LibTrueHD. Pinned to the minor for the same reason.
+        .package(url: "https://github.com/superuser404notfound/LibDTSX", .upToNextMinor(from: "1.0.1")),  // 1.0.1: rust_eh_personality weak (#752); dca 0.8.0 (harletty-bridge a717bae); the first two frames after a start or seek only prime the decoder
     ],
     targets: [
         .target(
@@ -64,7 +69,10 @@ let package = Package(
                 .product(name: "AetherFFmpegBuild", package: "FFmpegBuild"),
                 .product(name: "Dovi", package: "LibDovi"),
                 .product(name: "TrueHD", package: "LibTrueHD"),
+                .product(name: "DTSX", package: "LibDTSX"),
             ],
+            // Silent E-AC-3 Atmos that `AtmosOutputActivator` loops beside an APAC track on tvOS.
+            resources: [.copy("Resources/silent-atmos.mp4")],
             linkerSettings: [
                 .linkedFramework("AVFoundation"),
                 .linkedFramework("AVKit"),