Off-repo diagnosis (Pi 400, no RTC hardware, healthy crystal at -7.39 ppm):
the box's clock sat ~43 h behind and systemd-timesyncd's first contact after
the Oct 12 restart was 42.9 h later, so Pi-hole ran as a DNSSEC validator with
a wrong clock -> 1.69M+0.84M dnssec=3 (BOGUS) rows on 10-11/10-12 hitting the
PSN endpoints = the PS5's DNS errors. KSK-rollover theory retracted (both root
anchors present; root DNSKEY still returns ad). Pending creator decision:
tighten timesyncd polling and/or add a DS3231 RTC.
MyMistakes: the reply marker string (recovered verbatim, with the search
recipe), and the no-RTC/no-cron/timesyncd facts so the premise is not
re-derived a third time.
Planning-only session: no code, no build, no test run. The creator's vv
(9:16) webcam feature is now specified, and HANDOFF carries the locked
rules plus the code facts needed to build it without re-hunting:
- vv Live: small tile, circle OR rect (creator's choice), snapped to one
of six spots (top/bottom x L/C/R). vv Chat: rect only, plus enlarge to
fill one 607x540 area (exactly 50% of the frame), top or bottom.
16:9 and Starting/BRB/Ending unchanged.
- Placement by tap/click-drag with snap, like the socials, plus manual
entry in properties. vv keeps its own camera layout per scene, so the
16:9 layout survives switching modes (creator-confirmed).
- Geometry only: preview and compositor both read the element's own
X/Y/W/H + ClipShape, so the two-tile look falls out of the camera
covering half the band - no capture, chat-box or blit-path change. The
existing per-scene clamp already permits 607x540.
Withdraws the HANDOFF landmine "vertical-format regression ... no longer
shrinks the capture border in the preview": never verified, and wrong as
stated - the creator confirms vv dims the two bands astride the centre
band. It cost this session a pointless question.
MyMistakes.md records the session-history search recipe (the creator's
turns live in session_message, NOT part - a hit in part can be my own
text echoing the phrase) and the ask-don't-check lesson plus the
creator's standing reply-format directives.
The measurement unit existed to name which of the six scene elements eats ~140ms per
composite. The creator's Release recording (ty-20261016-1841-559-0002.mp4, 174.5s,
1920x1080@60) names none of them — the whole composite costs ~1ms:
composites 301 per 5s = 60/s, every slot, 0 dropped, 0 stalls after warm-up
avg render 1.0-1.1ms (worst 2.5ms) = ~7% of the 16.67ms budget
parts "#5 HD Pro Webcam C920 0.4ms, #5 ... raster 0.2ms, backdrop 0.2ms"
- the other four elements never reach the 0.1ms print floor
C4 cache 151R/150H: half the ticks were a whole-frame BlockCopy
webcam 30.0/s at the resolver, longest gap 50.0ms (one 60Hz frame of
sampling phase, not a drop)
the FILE 29.6 distinct webcam images/s, 82% of images held exactly 2 frames,
longest hold 267ms (the healthy 30fps-in-60fps signature)
The two stalls are not composite cost: 124ms on the very first render (warm-up) and
53ms entering the "Chat" scene (a rebake).
So the historic slow numbers are not a slow scene but different configurations:
- Oct 9 19:48 (only bad take still on disk) was the VERTICAL tier:
"FramePump started (1080x1920 @ 60 fps) scene='Live' comp=608x1080" - the 607x1080
centre crop upscaled to 2M px, 92 renders/5s, avg render 45.0ms, worst 278ms.
- Oct 5 12:01 (141ms) logged no scene=/comp= at all - the suffix did not exist yet,
so the map's "21:9 -> 1920x803" explanation stays a hypothesis, not a reading.
- Every slow window shows "cache xR/0H" (the whole-frame cache never reusing) while
this take shows 151R/150H, and one Oct 5 take latches mid-take (1.5ms -> 3.0ms ->
45.2ms in 15s, then stays). That latch is the lead for the next take; it is a
hypothesis and the parts line is what will decide it.
Corrections to the map (same commit, stale facts replaced not appended):
- "It was never Debug-vs-Release" is WITHDRAWN: same 1920x1080 tier, Oct 9 = 10.4ms
vs this Release take's 1.0ms, and Build <hash> carries no configuration - every
pre-2026-10-16 render number is config-unknown.
- The "same scene shape" 10:01/12:01 comparison is superseded: it was an eye
comparison of takes whose geometry was never logged.
New: scripts/webcam-cadence.py - the ground-truth probe for what the FILE holds,
beside what the pump says it did. It took three wrong readings to get right (H.264
re-encoding defeats exact-pixel comparison; a round-clip crop's corners leak 60fps
background; a fixed crop across a scene switch reads background), all recorded in
MyMistakes.md as a recipe, including an ffmpeg-native cross-check
(tblend=difference + signalstats.YAVG - the same metric, the technique already in
this repo's A/V-sync recipe) which agreed: 52.2% duplicates / 28.6 distinct/s vs
50.7% / 29.6/s on different windows of the same take.
Next: no fix is allowed yet - the creator makes a Release take in the failing
geometry (vertical 1080p60 tier first), and that take's parts line names the stage.
Docs in the same commit: ai.md (new "What the first instrumented take said" section,
the two corrections, the parts-line denominators, the tool in the inventory),
TASKS.md (reading outcome), HANDOFF.md (rewritten), MyMistakes.md (recipe).
Gates: clean build 0 warnings, full suite 428/428, scope check passed.
The suite failed Stats_ReportRealRendersAndDistinctWebcamFrames_NotJustPaddedSubmits
with "Collection was modified; enumeration operation may not execute" at its OWN
log.Any(...) line (FramePumpTests.cs:606): the test handed FramePump a plain
List<string> and enumerated it from the test thread while the pump reported stats
from its loop thread — pacingDelay is Task.Yield(), so the loop really is concurrent.
Production passes the thread-safe AppLog, so the app was never involved.
It passed 3/3 in isolation, which is exactly how a data race presents as a
load-sensitive flake. The trigger: the previous commit ("Per-element composite timing
+ scene dump") added log lines emitted FROM the pump thread, raising the collision
rate from rare to routine — a latent race in the old test, promoted by the new
traffic.
Fix: ConcurrentBag<string> (enumeration is a moment-in-time snapshot) for the sink,
2 call sites + the NewPump parameter. No app code touched.
Proof: full suite 428/428 passed, 0 warnings (the run that surfaced the failure was
before the fix; the same failure has not recurred).
Docs in the same commit: MyMistakes.md — "A 'known failure' label without a recorded
cause = a bug on life support" gained this second instance, with the rule that a
test failing ONLY under load means a shared mutable collection on the producer's
thread, not a retry.
The cadence work ended with a question the log could not answer: the compositor
manages 6.8 composites/s at ~141ms while the 10:03 take did 29-34ms on the "same
scene shape" (full-render split=0 elements=6 dynamic=5). Nothing in the stats line
said WHICH of the six elements cost what, and nothing recorded the scene itself, so
the two takes could only be compared by eye. This closes that gap and nothing else:
no fix is attempted until a real take names the culprit.
- Services/Compositor/CompositeTimings.cs (new): per-window collector.
Add(stage, ticks) accumulates; Describe(renders, minMs = 0.1) prints the averaged
breakdown slowest-first, dropping sub-0.1ms noise.
- SceneCompositor.Render / CompositeLayers / BakeStaticBase / BlitBakedElement /
BlitCachedLayer + SceneGraph.GetBakedBase: trailing optional CompositeTimings?
Every stage is timed (backdrop, base, flash, social, ticker, zoom, scale, overlay,
bake backdrop, bake overlays) and EVERY element gets its own entry keyed by
z-index + name ("#2 Webcam"), with the nested "raster" sub-stage recording the
paste-cache rasterize miss - the split between "re-samples every tick" and "slow
paste" without a second session. Bake-pass elements are prefixed "bake #..." so a
one-off rebake cannot masquerade as a per-composite cost. Callers that pass
nothing (thumbnails, existing tests) are unaffected and pay nothing.
- FramePump: owns the collector, threads it through RenderScene/RenderFull/
GetBakedBase/CompositeLayers, and logs "FramePump parts (N renders/5s): ..." once
per 5s window beside the stats line (averaged per render, so it compares directly
with "avg render"), then resets it. Also dumps the scene - "FramePump scene dump:"
header plus one line per element (z-index, name, kind, source type, visibility,
rect, opacity, clip) - at pump start and on every scene change.
- One Good Dog integration test (ytLive.Tests/CompositeTimingTests.cs): a real
composite of a full-canvas element plus a 64x64 one through the real collector
must rank the full-canvas element first, carry the raster sub-stage and the fixed
stages, and the dump must name both elements with z-index/type/rect and echo the
split point.
Derivative work (AGENTS.md): OBS's answer to "which source is slow" is the Source
Profiler - average execution time of each source's tick and render functions
(https://docs.obsproject.com/reference-libobs-util-source-profiler, util/profiler.h).
Same shape here: one timer per element per composite, reported as an average, plus
the stage split this case needs.
Docs in the same commit: ai.md (new subsection "Per-element composite timing + scene
dump", the closed gap corrected in "Frame cadence"), Services/index.md (new file row
+ FramePump/SceneCompositor rows), TASKS.md (unit marked shipped, next = creator
take), HANDOFF.md (rewritten: unpushed commit, next step, verify.sh whitelist hole).
Also recorded this session (from the creator): the launch command is
`dotnet run -c Release` - Release is the DEFAULT, Debug only when explicitly named,
and perf takes must read Release. ai.md -> Run updated.
Full suite 428 tests, 0 warnings; the only failures are the known audio env flakes
(both pass in isolation). Scope check passed.
The prior handoff claimed the strip-tile/on-air unit was uncommitted and the
pinned-chat unit was 'not mine - do NOT commit'. Reality: both are committed
(8e1d6f4, 931ab46) and origin/main is level with HEAD (0 ahead). Corrected to
match the code and recorded the true next step (webcam per-element composite
timing measurement). Removed the stray Assets/llamacasty-text-nobg-2.png~ backup.
(Docs: handoff)
- Model: Google's official MediaPipe Selfie Segmenter (landscape, Apache-2.0 per model card),
converted to ONNX by tools/selfie-seg/convert_tflite_to_onnx.py (custom op
Convolution2DTransposeBias -> ConvTranspose) and verified against MediaPipe's own runtime
(verify_against_mediapipe.py: max |diff| 5e-5, 100% binary agreement). Embedded resource, 430 KB.
- Services/BackgroundEffects: IPersonSegmenter seam, OnnxPersonSegmenter (onnxruntime 1.31.0 CPU,
never throws), BackgroundEffectProcessor (None passthrough, Remove = alpha mask). Not wired yet.
- Tests: Remove integration test (fake segmenter) + bundled-model contract test vs MediaPipe
reference mask. Full suite 407/408 — the one failure is the documented interactive-desktop flake
RealMouseDrag_OnTheLayerList_PersistsTheReorder (passed in the run before; code path untouched).
- Docs: ai.md reversal + section + licensing guard, THIRD-PARTY-NOTICES 11-12, TASK 51, indexes,
MyMistakes recipe, HANDOFF.
References: obs-backgroundremoval (approach only — ONNX Runtime on CPU, 2 threads; GPL-3.0, no code
used) https://github.com/royshil/obs-backgroundremoval ; MediaPipe image segmenter
https://ai.google.dev/edge/mediapipe/solutions/vision/image_segmenter
The zoom feature shipped without a task record while its index rows carried the
"TASK 51" label — which belongs to the webcam background effects (Remove + Blur).
For an honest catalogue and handoff:
- TASKS/task-52-zoom-box.md: feature provenance (creator-requested), the three
2026-10-08 corrections (loupe -> anchored capture over the full composite drawn
last -> never show the dim shading), shapes, implementation map, tests, DONE.
- TASKS.md: row 52 added (shipped); row 51 reworded — unblocked now that the
zoom build shipped.
- TASK 51 -> TASK 52 in Models/index.md, ViewModels/index.md, and the two code
comments that pointed at the zoom as "TASK 51".
- HANDOFF.md: pre-push readiness notes (5 commits ahead, push-safe).
Creator's final zoom correction: as the box is dragged, its capture region slides
under the dim rectangles and the box picks up the 55% black "div shading" inside
itself, because the VisualBrush samples the whole CanvasGrid and DimRects lived
inside CanvasGrid. The captured area must stay clear regardless of box position.
- PreviewPane.xaml: the DimRects ItemsControl is now a SIBLING of CanvasGrid
(above the composed content, below the ZoomBox), never a descendant — the zoom
source is shading-free by construction.
- ZoomBoxInteractionTests: structural invariant — DimRectsLayer is a direct
child of PreviewCanvas (not inside CanvasGrid) and renders below the ZoomBox.
- Side effect (correctness, not a bug): quality dims now also sedate the global
overlays outside the crop, matching "chrome outside the crop is dimmed".
Creator corrected the "loupe" zoom on 2026-10-08: the click anchors a fixed
capture region (box/2 centred on the click) and the box is a live 2x view of
that region — dragging the box never changes its content; it magnifies the
fully-composited frame (every layer, incl. webcam) and draws LAST (n+1).
Rectangle shape is 4:3 ("more of a 640x480 aspect").
- MainViewModel.Zoom.cs rewritten: ActivateZoom anchors CaptureX/CaptureY
(independently clamped); UpdateZoomMove relocates the box only; the per-frame
BuildZoomFrame/ZoomFrameBuilder path is gone.
- SceneCompositor: skips the active ZoomSceneElement in every layer loop and
BlitZooms it last via the new pure ZoomMagnifier (2x bilinear, temp snapshot
via ArrayPool, round clip).
- Preview: ZoomBox hoisted above CanvasGrid as a sibling; its brush samples the
whole composed preview at the fixed capture (2x).
- Tests: ZoomMagnifierTests (box-over-blue shows green capture at 2x incl. an
extra red "layer" — the loupe/backdrop-only regressions both fail it) +
round-clip; ZoomBoxInteractionTests rewritten (anchored capture, move-doesn't-
change-capture, 4:3/9:16/square). 404 pass, 2 known environment flakes.
Right-click now creates a transient ZoomSceneElement at z-index 1 whose pixels are a
genuine 2x magnified sample of the backdrop (live capture under art), composited into
stream/recording like any layer while the output crop stays the tier. All shapes are
centered on the click and clamped to frame; healed on load/scene-create, never persisted,
hidden from the layers list + element template (IsZoomLayer). BuildCompositorOptions is
now internal = always tier crop (test seam). ZoomFrameBuilder is a pure static sampler
(box/2 window centered on the box center, single-slot cache) proven by a banded-backdrop
test; LayerReorderPersistenceTests drag resolves rows by item since the collapsed zoom
row shifts container indices. Also books TASK 51 (webcam background Remove + Blur) and
syncs memory docs (HANDOFF, MyMistakes, indexes). Full suite: 405 passed, 0 warnings.
Right-click in the live preview now procs a box that is actually visible and
populated with the zoomed region, and each Zoom View Shape sculpts the box per
spec. OBS-style derivative: magnifier box shows the composited region and dims
the surroundings (same pattern as OBS "zoom" / preview region highlights).
Root causes of the invisible box (two WPF traps, recorded in MyMistakes.md):
- Empty <VisualBrush Stretch="None"/> used as an OpacityMask paints nothing,
so the mask alpha is 0 and the WHOLE box (border included) went transparent.
- ZoomBox lived inside CanvasGrid while its VisualBrush sampled CanvasGrid — a
circular reference WPF renders as nothing. Hoisted ZoomBox to a sibling of
CanvasGrid inside a new PreviewCanvas wrapper (same 1920x1080 space).
Shape semantics (MainViewModel.ActivateZoom):
- Circle: square, click point = center.
- VerticalRectangle (phone): tall 9:16, click point = top-left corner.
- HorizontalRectangle (tablet): wide 16:9, click point = top-left corner.
- Boxes stay fully inside 1920x1080 via Math.Clamp.
UI: ZoomRectOutline/ZoomCircleOutline shapes replace the square Border;
circle clips the box to an EllipseGeometry; brush Visual released when inactive.
Zoom View Shape dropdown now uses the standard YtComboBox style.
Test (Good Dog rule, one integration test): ZoomBoxInteractionTests verifies
aspect/anchoring/clamp per shape and the matching outline+clip in the real
MainWindow. Full suite 404 passed, 0 warnings.
Memory: HANDOFF.md zoom state updated; WPF VisualBrush traps appended to
MyMistakes.md.
A composite splits into four horizontal row strips across four DEDICATED
background threads; the calling thread takes a strip too (so four workers
never mean five-way oversubscription), and anything under 64 rows runs
inline. BlitContentRaw, BlitBakedElement and BlitOverlay now fan out.
Output is byte-identical — strips already partitioned by row.
The demotion deliberately does NOT come from Parallel.For. The pool is
shared, so Thread.CurrentThread.Priority = BelowNormal inside it leaks
into every unrelated work item that thread later runs — a full test run
after trying it broke AudioPipelineTests.Mix_HonorsProviderGains_AndGameMute_
KillsTheLoopback. Established guidance agrees: use your own producer/consumer
queue with your own Threads and set their priority yourself
(https://stackoverflow.com/questions/524483/net-backgroundworker-objects-thread-priority,
https://stackoverflow.com/questions/1578599/c-sharp-lower-thread-priority-in-thread-pool).
Helpers/StripPool.cs is that queue. Four threads is the measured sweet
spot: fewer leaves render over budget, more steals cycles from the game
and the encoder and made render worse.
The scaled blit's column LUT moves from a shared static buffer to
ArrayPool<float>.Shared with a finally — a single static buffer is a data
race the moment strips run in parallel.
Gate: SceneCompositorTests.StripParallel_RepeatedRenders_AreByteIdentical —
8 renders of a scaled round webcam + mirrored image + master-sized flash over
a short backdrop must be byte-for-byte identical, and must differ from a
backdrop-only render so it cannot pass vacuously. It caught a loop-variable
capture (new Thread(() => Worker(i)) saw i == Workers) that crashed the test
host on the first run.
Handle the case where an opaque source exactly covers the destination
canvas with 1:1 dimensions (even if offsets differ in calculation). This
avoids per-pixel alpha branching and mirrors the existing memcpy fast
path. Byte-identical for opaque sources.
Add timing for social bar, ticker, flash delegates and scratch acquire,
and extend the 5s stats line to report them (ms per render). No behavior
change, only measurement.
The Stream Alerts properties panel was incorrectly showing for the webcam
layer (a WebcamSceneConfig is not a Source) because IsAlertBox only existed
on Source, causing WPF binding to be treated as false-but-invisible context
or the trigger to bind against the wrong type. Adding a virtual IsAlertBox
to the base SceneElement (default false) and overriding on Source prevents
the UI block from rendering under webcam selection. Scope: Models/SceneElement.cs, Models/Source.cs, ytLive.Tests/AlertBoxPreviewVisibilityTests.cs. Build/test: app + 4 alert-box tests pass, other failures unrelated (float precision).
Take ty-20261005-1201-556-0001.mp4 (75.3s) plus its log settle it. The
C920 delivers 30.0/s with ZERO gaps over 50ms and its own device stamps
agree - camera, USB and driver are clean. The pump managed 34 composites per
5s at ~141ms each, and the file holds 6.8 distinct webcam images per second
at 88.7% duplicates with no arrival hitch whatsoever. The judder is us
holding one face for 140ms and padding the rest of a 60fps container, which
is precisely what a lipreader experiences as "cutting out".
So 720p is a SHARPNESS win, not a smoothness win: it triples the bytes the
compositor stretches per render on a path already 4.4x over budget. The
exclusive camera lock stays agreed and stays queued, but it moves after the
compositor work or it makes the judder measurably worse.
The number still unexplained: the 10:01 take hit 25.6 images/s and this one
6.8, with the same scene shape (full-render split=0 elements=6 dynamic=5)
and the same 640x480 camera - composites went 29-34ms to 143-155ms between
sessions and nothing logged says why. There is no per-element composite
breakdown and no scene dump at pump start, so the next unit is measurement,
not a fix. Recorded the pump-counter overcount (13/s vs 6.8/s: one 141ms
composite lets ~4 camera frames arrive, and the counter stamps the first
epoch the resolver sees, not the frame actually drawn) so it is not
re-derived as a bug later.
gramps