gramps/LlamaCasty
0
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

docs: update HANDOFF + MyMistakes.md — take-21 web transparency fix landed

Browse Source
This commit is contained in:
Llama Chile Shop
2026-09-06 12:58:05 -07:00
parent e002847777
commit 5ca9778e23
2 changed files with 36 additions and 157 deletions
Download Patch File Download Diff File Expand all files Collapse all files
HANDOFF.md
+21 -144
View File
@@ -1,157 +1,34 @@
# HANDOFF — Session State # HANDOFF — 2026-09-06
## Branch / Commit State ## Branch / Commit State
**`main` HEAD = `6d11e8b`** — `#12` landed (per-element raster reuse: _elementRasters, steady-state compositor allocation ≈ 0; regression test; 289 tests). **`#11` landed as `d35823a`** (release counter in wordmark; capture/camera/web rings 4→8 + Epoch — three-loop hunks dropped per plan). User tested take-15 frame (~clean). Recording output: webcam "black square" regression under the web overlay is OPEN (spin guard triggered; research first before next attempt). **`main` HEAD = `e002847`** — web transparency fix (take-21): feed full canvas to compositor.
**`today-slices` branch @ `ce87490` (#17)** — preserves ALL 17 commits of the 7fcb2ad→ce87490 chain ```
(1 `716a77f` … 17 `ce87490`), including the two-loop producer, encoder staging fix, chat fix, e2e e002847 fix(web): feed full canvas to compositor — transparent margins reveal layers beneath (take-21)
harness. Nothing is lost; never hard-reset away from it without preserving. 70db344 perf(camera): MJPEG negotiation before frame reader — camera drives at its best rate
0f059ec docs: HANDOFF — #12 landed (element rasters), MJPEG next, web transparency spin-guarded
6d11e8b perf(compositor): reuse per-element rasters in-place — take-16 GC-churn fix
87509bc (origin/main) docs: restructure TASKS.md into a catalog — one file per task in TASKS/
```
## ROLLBACK RECORD (2026-09-04) **Pushed to origin: `87509bc`.** 4 unpushed commits after it.
- **Rolled back to #9 `c46f3a1`** via `git reset --hard` on `main` (user-directed) after the user **`today-slices` branch @ `ce87490`** — preserves 17-commit chain. Fallback recovery.
confirmed #10 introduced the tearing. Rolling forward WITHOUT #10 (per user directive) starts from
this commit.
- **`today-slices` @ `ce87490`** preserves the entire 17-commit chain — the fallback recovery point.
## REJECTED this session (2026-09-05) — web transparency fix FAILED, reverted ## Resolved this session
**`0f441d8` "composite consumes the FULL canvas" was wrong.** Theory: the alpha-crop fed to the - **Audio regression (TASK 26):** Non-event — external headset volume control was at minimum, not a code bug.
compositor's UniformToFill zoomed opaque content over the whole element rect → black box. Fix: hand - **Webcam black square / web overlay transparency:** FIXED at `e002847`. Root cause: `bccdb48` introduced alpha cropping; the CROP was fed to the compositor, which `UniformToFill`-zoomed the opaque content to cover the entire element rect. Fix: feed the FULL `1920×1080` canvas to the compositor; transparent margins reveal layers beneath; the alpha crop (`outPixels`) is used only for the preview `WriteableBitmap`. `0f441d8` (take-19) had the right insight but introduced `CanvasRing` alongside the fix — the ring reuse caused additional breakage. Take-21 keeps the single `CanvasScratch` buffer (no ring), feeds full canvas, Epoch increments per frame.
the compositor the full transparent-background canvas (8-deep ring + Epoch), keep the alpha crop for
the preview/selection only. User tested: **"didn't work at all, and broke additional crap."** Reverted
as `1295e0e`. Tree verified back at the #11 state; builds 0 warnings both projects.
Two hard facts the theory did NOT anticipate: ## MyMistakes.md update needed
1. An opaque dark band still covers the webcam region in the RECORDING (the "broke additional crap"
may be the widget reappearing over the webcam in the preview/preview sizing — unconfirmed).
2. The preview half looked fine while the recording half was dark, per the user, all along.
**SPIN GUARD TRIGGERED (AGENTS.md Working rules):** this is the second failed remedy for the same The spin-guard entry for the web overlay needs updating with the take-21 resolution. See that file.
symptom (web sources composite dark/opaque over other layers in the recording). Third-guessing over
the same code is forbidden. Next attempt MUST start with external research — how OBS / CEV /
other overlay toolchains keep browser-source transparency in the *recorded* output, cited in the
commit message AND MyMistakes.md before coding. Likely seams to revisit after that research:
`CapturePreviewAsync` PNG alpha → `FormatConvertedBitmap` w/ `PixelFormats.Bgra32` (does the
premultiplied/straight-alpha conversion drop alpha?), and the preview-vs-recording consumer
difference (preview = WriteableBitmap on the dispatcher; recording = composite read on the pump
thread) — but NO code until the research says what to do. Also: the user's very first observation —
"webcam blacked out by the animated web resource layer" — may mean the widget's HOST page itself
paints an opaque/dark background (injection script wouldn't fix a page that sets its own bg) — verify
against a blank page, not an alert host, before blaming the compositor again.
## The commit list (for picking this up anytime)
Numbered oldest→newest (#0 = the working baseline the user picked: `7fcb2ad` "zero known failures era
begins"). **#10 is the breaking commit** (two-loop producer = the recorded-output tearing). #9 is the
last working commit. \(SHAs = full 40-char hashes below\):
| # | SHA | Subject |
|---|-----|---------|
| 0 | `7fcb2ad…` | baseline: "zero known failures era begins" (last build the user trusted before the chain) |
| 1 | `716a77f` | take-3 starvation fix (deadline pacing + row-blit) |
| 2 | `432adfd` | slice 2 — integer bilinear + pump scratch pool |
| 3 | `1c48849` | slice 3 — chat raster cache |
| 4 | `27bf743` | GUID build stamp |
| 5 | `6af2026` | slice 5 — paste cache |
| 6 | `09a866e` | slice 6 — sleep quantum fix (timeBeginPeriod + spin tail) |
| 7 | `eb4c379` | slice 7 — pump off UI thread |
| 8 | `fbc8562` | slice 8 — screen-capture buffer ring + Epoch + gen2 stat |
| 9 | `c46f3a1` | **docs — LAST WORKING (tree = slice 8); main sits here now** |
| 10 | `8260474` | **⚠ BREAKS: two-loop producer — introduced the output tearing; never lands** |
| 11 | `6fd1d9c` | #11 — release counter in wordmark + capture/camera/web rings 4→8 (FramePump hunk = two-loop-only, drop) |
| 12 | `d1126dd` | #12 — compositor per-element raster reuse + MJPEG camera negotiation (FramePump hunk = two-loop-only, drop) |
| 13 | `e4621c1` | #13 — docs (HANDOFF rewrite) |
| 14 | `b4f6a92` | #14 — docs (recording audit) |
| 15 | `7bf00a9` | #15 — e2e producer-smoothness bar (thresholds tuned to two-loop; re-tune on single loop) |
| 16 | `1a808ae` | #16 — chat fix: empty-buffer chat box composites |
| 17 | `ce87490` | #17 — encoder staging snapshot (the genuine tear cure; belt-and-suspenders on the single loop) |
Full SHAs: 0=`7fcb2ad…`(baseline) · 1=`716a77f` · 2=`432adfd` · 3=`1c48849` · 4=`27bf743` · 5=`6af2026`
· 6=`09a866e` · 7=`eb4c379` · 8=`fbc8562` · 9=`c46f3a1` · 10=`8260474` · 11=`6fd1d9c` · 12=`d1126dd`
· 13=`e4621c1` · 14=`b4f6a92` · 15=`7bf00a9` · 16=`1a808ae` · 17=`ce87490`.
Roll-forward plan (user's method, ONE at a time, user tests each): apply #11 → #17, skipping #10
entirely; drop each commit's two-loop-only FramePump hunks.
## What just happened (2026-09-04, bisect verdict)
The user bisected the 17-commit chain by directing resets of `main` + builds + live runs. **Verdict:**
- **#10 `8260474` (slice 9, the two-loop producer) INTRODUCED the screen tearing over the layered
elements in the recorded output file.** #9 (`c46f3a1`) does NOT have it. Tomm every time.
- **Mechanism (closed):** RenderLoopAsync owns a private 4-deep crop ring and publishes the newest
composite; OutputLoopAsync calls `SubmitFrameAsync(published.Frame)` — and at #10 the encoder still
writes `frame.BgraPixels` DIRECTLY into the async stdin pipe (no copy). When the encoder/pipe stalls
>~50ms, the free-running render loop laps the 4-deep ring and REPAINTS the array being drained →
the recorded frame is a mix of two composites = the tearing. Preview is immune (synchronous
WriteableBitmap copy). The #10 code comment admits it: "torn frame if the renderer laps the ring
during a >~50ms encoder stall — bounded cosmetic risk". #9's single serial loop cannot tear by
construction (same thread owns render→submit→release).
- The genuine fix for the tear is `Service c87490`'s core: snapshot `frame.BgraPixels` into an
encoder-owned staging buffer inside `SubmitFrameAsync` BEFORE the async write. It was bundled into
#17 and rejected wholesale with it.
**DECISION (user directive):** roll forward from #9 WITHOUT landing #10's two-loop producer.
**User's method:** apply #11→#17 one at a time; FIRST tell him what the commit does/fixes; HE tests;
we proceed or drop. Do exactly what is asked, no more, no suggestions, until the code is "unfucked".
## Verified contents of #10 (for the roll-forward)
**Unsafe (the tear, never lands):** `Services/Encoder/FramePump.cs` — the whole two-loop rewrite.
**Safe, producer-side allocation only (no render/submit-path contact):**
- `Services/Audio/AudioMixer.cs` — `MasterOutputGain = 1.4f` pre-limiter (+40% creator ask; paired
`AudioPipelineTests` bound 0.55→0.7).
- `Services/MediaCaptureFrameSource.cs` — 4-deep camera ring + `Epoch` (ends ~110-220MB/s LOH churn).
- `Services/WebView2Manager.cs` — reused canvas scratch + 3-deep out-ring + reused WriteableBitmap
(ends two fresh arrays + a new bitmap per capture tick).
## What was attempted / aborted this session
Cherry-pick of `6fd1d9c` (#11) onto #9 was started, hit conflicts, and was ABORTED on the user's order.
Cleanup done: `cherry-pick --abort` fails for `-n` (no sequencer state) → `git reset --hard HEAD`.
Tree verified clean at #9. **Learned (record before re-applying):** #11 on the #9 base = release
counter (#11's "non-ring half", [TopBar.xaml, ytLive.csproj, BuildStampTests.cs]) + ScreenCapture ring
4→8 + NEW camera ring + NEW WebView2 pooling (these blocks are ADDED fresh in #11 at depth 8, not just
widened) + docs. #11's FramePump hunk (`ring = new byte[8][]`) is two-loop-only → DROP. #12's FramePump
hunk (stats formatting) is two-loop-only (`TryReportStats`) → DROP. #13/#14 are docs-only. #15 e2e bar
was thresholded against two-loop measurements. #16 (chat) and #17 (encoder staging) are pump-independent
and apply cleanly.
## OPEN — next, in order
0. **Webcam-black-square-under-web transparent: research FIRST.** Second failed remedy → spin guard.
Externally research how OBS/CEV keep browser-source transparency in recorded output; cite the URL
in the commit AND MyMistakes before any code (see REJECTED section above). Do NOT touch the
compositor/web code until research names a seam.
1. **#12 MJPEG camera negotiation** — `MediaCaptureFrameSource` MJPEG negotiation before frame reader
creation (SetMediaStreamPropertiesAsync → best MJPEG ≥640×360 @≥30fps, ≤1280 wide), startup.log
truth line of actual media type. Stat formatting (`:F1` ms) already in #12 diff. Next to apply.
2. **#13 `e4621c1`**, **#14 `b4f6a92`** — docs (TASKS restructure this session).
3. **#15 `7bf00a9`** — e2e smoothness bar (re-tune thresholds to single-loop measurements if needed).
4. **#16 `1a808ae`** — chat empty-buffer composite fix + test.
5. **#17 `ce87490`** — encoder staging snapshot (the genuine tear cure — applies on the single loop as
belt-and-suspenders).
6. **Unit B (top bar + session logic)** — spec + settled decisions are in the old HANDOFF history on
`today-slices`; don't re-ask.
## Landmines ## Landmines
- testhost shares startup.log with the app — filter by time when triaging. - testhost shares startup.log with app — filter by time when triaging
- Stale testhost/exe locks the DLL (MSB3027): `taskkill /F /IM testhost.exe` / `ytLive.exe` first. - Stale testhost/exe locks DLL: `taskkill /F /IM testhost.exe /IM ytLive.exe` first
- Do NOT run full-suite vstest (WASAPI hang, pre-existing); flow = clean build + per-class + scope-check. - Do NOT run full-suite vstest (WASAPI hang); flow = clean build + per-class + scope-check
- Pacing-seam fakes MUST await/yield (sync-completed task → pump runs inline on StartAsync → hang). - Kill app before build: `/mnt/c/Windows/System32/taskkill.exe /F /IM ytLive.exe`
- `FakeEncoder` snapshots submitted frames — keep any new fake encoder honest about buffer recycling. - ffmpeg month-end pinned; `Startup.log` "Recording saved:" shows real final path
- Multi-stage fixed-point: shift only at the end (MyMistakes 2026-09-04).
- Real-`MainWindow` tests: `LayoutPathOverride` + temp DB mandatory; `VolumePushOverride` for volume.
- ffmpeg: month-end pinned build; `Startup.log` "Recording saved:" lines show the real final path.
- Kill the app (`/mnt/c/Windows/System32/taskkill.exe /F /IM ytLive.exe`) before any build that
replaces the running binary.
## @ User note
**Do EXACTLY what the user asks, no more, no suggestions.** He is driving the unfuck personally:
commit-by-commit roll-forward, he tests each, proceed or drop. Keep responses SHORT. He is high-voltage —
do not spin, do not theorize, do not touch docs/commits he didn't order. Commit every work unit; push on
his word only.
MyMistakes.md
+15 -13
View File
@@ -15,20 +15,22 @@
## 🔬 Recipes registry ## 🔬 Recipes registry
### ⚠ FAILED remedy (spin guard, 2nd strike) — web overlay composites dark over the webcam ### ⚠ SPIN GUARD TRIGGERED → RESOLVED (take-21) — web overlay composites dark over the webcam
(2026-09-05, takes 19-2/20 + user re-test) Attempt `0f441d8`: feed the compositor the FULL **Take-19 (`0f441d8`): FAILED.** Attempted to feed the compositor the FULL canvas instead of the
transparent-background canvas instead of the alpha crop (crop = measure-only). User verdict: alpha crop. User verdict: "didn't work at all — broke additional crap." Reverted (`1295e0e`).
**"didn't work at all — broke additional crap."** Reverted (`1295e0e`). I hypothesized the The crop→UniformToFill zoom insight was RIGHT, but the commit ALSO introduced `CanvasRing` (8-deep
alpha-crop→UniformToFill zoom; the fix changed that and the recording was still dark, so the full-canvas buffer ring) alongside the fix — the ring reuse poisoned the paste cache (array identity
crop→zoom was NOT the mechanism. **Due to AGENTS.md's spin guard this is the second failed remedy — collisions via the Epoch mechanism), causing additional breakage beyond the transparency issue.
NO third guess over the same code.** Before the next attempt: research externally how OBS/CEV keep
browser-source transparency in the *recorded* composite (cite URL), AND test the simplest falsifier **Take-21 (`e002847`): FIXED.** Root cause confirmed: `bccdb48` introduced `FindContentBounds`
first — a blank transparent page with no widget content — to decide whether the dark band is the alpha cropping; the CROP was fed to the compositor (`session.LatestFrame = VideoFrame(cropW, cropH,
widget's own page background (host pages that set their own bg defeat the injection script) rather outPixels)`). `UniformToFill` then zoomed the opaque crop content to cover the whole element rect,
than the compositor. Do not re-derive; the seam candidates (CapturePreviewAsync PNG alpha through covering whatever was beneath (webcam). Fix: keep the single `CanvasScratch` buffer (no ring), feed
FormatConvertedBitmap premultiplied/straight; preview-dispatcher vs pump-thread consumers) stay the FULL `1920×1080` canvas to the compositor (`session.LatestFrame = VideoFrame(pixW, pixH,
unconfirmed and untouched until the research names one of them. pixels)`), use the alpha crop only for the preview `WriteableBitmap`. Epoch increments per frame so
array identity stays unique without the ring. The `0f441d8` insight was correct — the ring was the
"additional crap."
### Shrink / re-encode an image for the README (screenshots → small hero image) ### Shrink / re-encode an image for the README (screenshots → small hero image)