LlamaCasty/HANDOFF.md

HANDOFF — 2026-11-12 (bitrate correction, 4 units; unit 1 landed)

State

• main; b265d42 is the latest pushed commit. 1e3a4f7 and everything after are local.
• Done: TASK 35 scene-linked audio, TASK 55 disk hygiene, TASK 65 replay clips, TASK 37 defaults/current split.
• Recovery fixes: step-down closes/recreates audio in order; recovery notifications/stops marshal to the dispatcher.
• App Settings: six toast positions; sticky/click-to-dismiss alerts (off by default preserves timers).
• 1e3a4f7: OutputProfile.EntitlementEnforced = false — Release builds can select 60 fps pre-release, and the licence indicator row is suppressed while nothing is limited. Flip at launch.
• Creator test was stable (~31 fps, 0.2% drop) but the stream's own readout averaged ~5.1 Mbps against an @8 target: the args pass -b:v/-maxrate/-bufsize with no -rc, so h264_nvenc ran its default VBR and undershot. Not a machine limit (1.8 ms render, 0 stalls) and not a YouTube step-down (no step-down logged). Diagnosis is args, not ingest.
• Verification: 667 tests green, clean build with 0 warnings (./scripts/verify.sh).

Next action

• Unit 2 of 4: explicit CBR in Services/Encoder/FfmpegArgs.cs, spelled per encoder (NVENC/AMF -rc cbr; QSV and OpenH264 differ — verify, do not guess). One test/commit.
• Unit 3: rate presets + free-form Mbps per tier (min/rec/max from YouTube's H.264 table; default 12 at 1080p60, 10 at 1080p30, 6 at 720p), tier minimum becomes the step-down floor.
• Unit 4: local per-tier history; propose a rate change between sessions only, ≥10% above a zero-step high-water mark across 5 sessions, or repeated step-downs downward; explicit consent, "keep mine" remembered.
• Then resume 40·A/C/D → Move → 9·5 → 14·16/17/18 → 66 → 67 → 68 → 69 → 70 → 36.
• Do not push app commits absent a checkpoint request.

Landmines

• Assets/starting-background.jpg is modified in the working tree, awaiting the creator's visual check — do not commit it as part of another unit.
• ChromaKeyRealClipTests.RealClip_PlaysAtItsNativeRate failed once in a full run and passed alone: a load-sensitive flake, not the known pre-existing pair. Watch it; do not chase it as a regression.
• Early step-down path reuses ChangeEncoderAsync, so every rate change is a full encoder relaunch.
• Preview frame-source layers need a PreviewPane DataTrigger; pin each new type with a visibility test.
• Media sessions are keyed by path + look; never acquire by raw path.
• Audio pipe allows one server per name; DEBUG suffixes it, Release does not.
• Any test constructing MainWindow must use a temporary LayoutPathOverride and clear it in finally.
• Build with the Windows dotnet.exe host from WSL; read render numbers in Release.