BOSS / Aether matched-media Mac comparison — 4 November 2026

Read summary.json before using timings. Four HEVC-in-TS HLS BOSS runs have
confirmed missing video and are excluded from valid video comparisons.
The initial 84 runs measure clock progress, not per-run visible rendering.

Hardware: Apple M6 Mac mini / 12 CPU cores / 24 GB / macOS 27.0.1 (26A434),
arm64, Swift 6.4, AC power. Both builds use SwiftPM Release optimization.

Upstream AetherEngine commit:
a6fbf0a1c469df98df66d886482b9f062cdab7f8
https://github.com/superuser404notfound/AetherEngine
Default optional traits were used; DualLayerDolbyVision / TrueHD are disabled.

Media: use the exact upstream recipes at that commit with the recorded
FFmpeg 8.1.3-Jellyfin encoder. No scaling or encoding changes were made.
Scripts/fetch-fixtures.sh provides the main corpus.
Scripts/generate-avi-mpeg4-cache-fixture.py provides the AVI cases.
Scripts/make-hdr10plus-fixture.py provides the 2-frame / 0.2 s HDR10+ case.

The supplied fixture-retention scripts preserve the VOD and long-GOP files
from the original upstream scripts before temporary-file cleanup, replacing
only the final swift-test invocation with a copy. Put these beside the
original Scripts files. Encoder commands are unchanged.
Set BOSS_BENCH_FIXTURES to the absolute existing Fixtures directory.
Generated byte hashes can differ with other encoder builds; always feed
the same generated byte files to both players and retain their SHA manifest.
No historical private author encodes or FEL clips are included.

Harness:
- BossBench.swift uses BOSSPlayerCore and BOSSPlayerUI from the matching SDK.
- AetherBench.swift uses AetherEngine from the pinned upstream commit.
- Both are macOS executable targets with a mounted 640x360 AppKit/SwiftUI
  window. Build each target in Release and record source/dependency hashes.
- No paid SDK implementation or binary dependency is distributed here.
- The final BOSS adapter has additive VLC/native-track/HDR diagnostics;
  these extra fields were not present in the initial 84-run matrix.

Export the following environment variables:
  BOSS_BENCH_FIXTURES — directory containing all upstream-generated files
  BOSS_BENCH_OUTPUT — new output directory (defaults to ./benchmark-output)
  BOSS_BENCH_BINARY — absolute path to built BossBench executable
  AETHER_BENCH_BINARY — absolute path to built AetherBench executable
Then run sequentially:
  python3 run_matrix.py
  python3 run_network.py

Both scripts start fresh processes; they alternate engine order by repetition.
The second script owns a local HTTP range/paced origin and stops it on exit.
It opens no provider credentials or private network media.
Do not run other playback tests concurrently while collecting timings.

These are portable public-API adaptations using upstream media. They do not
execute every original Aether internal regression assertion. Read the
upstreamProtocolLimits field in summary.json. Timings include clock advance
to 0.5 s and are not first visible frame latency, energy or system-wide RSS.

Missing exact FEL media:
fel-clip-a.mkv, fel-clip-b.mkv, fel-clip-c.mp4.
Obtain originals and their hashes before claiming an exact FEL comparison.

Final rendering audit and corrections:
The second 84-run paired audit includes native-layer and VLC display witnesses.
It found 3 BOSS missing-video AVC cases (same fixture, 3 repeats). Preserve these
raw failures; do not relabel them as successes. The separate corrected AVC
acceptance records 3/3 rendered repeats and native/AVI/HLS controls.
After building the corrected SDK/harness, the same environment variables drive:
  python3 run_fix_acceptance.py
  python3 run_avc_acceptance.py
  python3 run_long_hls.py
Use a fresh BOSS_BENCH_OUTPUT directory for each script. The latter checks
the original generated 20-minute HEVC HLS file; BOSS seeks 12/90/4 and Aether's
harness seeks 12/4. Only equal targets may be compared.
final-comparison.json excludes failed AVC cases from paired aggregate timings.
No paid SDK source or correction patch is included in this public bundle.
