13.19g: a raster commit's continuation runs where the commit ends, as Swift's await returns; nothing lands between

TASKS.md

@@ -467,7 +467,7 @@ Reviews for this section ran on codex `gpt-6.1-sol` at `max` (agent `codex-sol`)
   - 13.13 and 13.16 (`70318de`): a typed Move-bar value then a real canvas press (the field finishes from the event loop on `MouseFocusReason`; `MoveBarPressTests`), the folder case, Alt-centre and long-drag resize snapping and the edge tie (`ResizeSnapTests`, `ResizeSnapCanvasTests`, fixture `MovePressFixtures.h`), the dead one-pixel Liquify "fix" removed, a Liquify dab's scratch allocation caught (skips the dab with a warning), the Radius field's `inf`, `WarpStroke::committedPoints` with its test, `softened`'s mask flag removed, and AGENTS and TASKS lines.
 - Known failures: at 81894f7 the full suite under `ctest -j16` failed `ColorRangeTests`, `GuideGridTests`, `GridSettingsSheetTests`, `TypeFontPreviewTests` and `GuideCanvasTests`, each of which passes alone: some state is shared between test processes (suspect `QSettings` or `ToolDefaults` keys). Find and isolate it; never loosen a check.
 - Next, in order: build; run the commits above and fix them; run the full suite and fix the parallel failures; review 3 over every fix since 81894f7 (13.8 included); then this item: acceptance run, distro checks (arch, fedora, fedora44, nixos, resolute), the `v1.4.5` tag and release, live installs, and the acceptance checklist. The 1.4.5 version, change log, README and acceptance table (results pending) are already in.
-- Resumed 2026-10-03 after the reinstall. The user's rule of that day replaces the ten-review allowance: every step below is reviewed by codex (GPT-6.1-Sol, max) in a herdr pane before the next, at most three rounds a step. Steps, in order:
+- Resumed 2026-10-03 after the reinstall. The user's rule of that day replaces the ten-review allowance: every step below is reviewed by codex (GPT-6.1-Sol, max effort until 13.19g, high from then at the user's word) in a herdr pane before the next, at most three rounds a step. Steps, in order:
   - [x] 13.19a The spec checkout is `Compositor-Ref` inside the repository (ignored by git, left out of the flake's source), at `v1.4.5`; AGENTS.md and TASKS.md say so.
   - 13.19a review: round 1 FAIL (the flake copied the spec; no step for review 3's scope; the budget unstated; the status said done), fixed; the Docker context finding answered (no Dockerfile copies from it). Round 2 PASS.
   - [x] 13.19b Build main in a fresh `omaphoto-dev` image.
@@ -488,7 +488,10 @@ Reviews for this section ran on codex `gpt-6.1-sol` at `max` (agent `codex-sol`)
   - 13.19f work: the three commits' tests pass. Mutants on their code: the tab strip's release watch (its button test, its start, its ending, its other-tab guard) and the glow's whole steps. The other-tab guard (Swift's `guard reorder?.id == id`) survived; `anotherTabPressedBeforeTheWatchMovesNothing` now catches it. The glow's rounded-up small sides are equivalent (the padded sides are whole steps).
   - 13.19f review: round 1 FAIL: `aDoubleClickInTheStripMovesADivider` pressed twice at one place, so a double click that skipped the divider passed. Now the first click and the double click land apart; that mutant is caught.
   - 13.19f round 2: PASS.
-  - [ ] 13.19g The full suite under `ctest -j16`; the shared state behind the five parallel failures found and isolated.
+  - [x] 13.19g The full suite under `ctest -j16`; the shared state behind the five parallel failures found and isolated.
+  - 13.19g work: eight full runs at `-j16` and `-j32` on 32 cores passed 330 of 330; the five failures of `81894f7` did not come back, alone or with the nineteen suites that share `QStandardPaths`' test folder, so their cause is not known (the old runs had a 12-core cap and bwrap). Under a 4-CPU cap at `-j16` a new one failed three of three: `AcceptanceTests::paintingTools`' rectangle made no layer. Diagnostics showed the gradient edit still pending when the drag began, with the project free: `finishRasterCommit` freed the project, then posted the continuation that clears the gradient, so for one turn of the event loop the edit refused every press (Swift's caller resumes from `await` in the same job). The continuation now runs where the commit ends, for the gradient, the pixel move and fills; their callers' callbacks stay posted. Round 1 of review found the same gap one level up, through posted callbacks that carry work: the floating selection installed after the clear, and a filter, Levels, Hue/Saturation or Invert begun after a pending gradient read whatever layer was active by then. With the continuation inline, both run where the commit ends; only early paths post. `CommitContinuationTests` posts a check as the project frees: the gradient gone, the move gone, the floating edit ready, a tool chosen then staying chosen, and each chained command keeping its layer; the two first were red before the fix, posting the continuation again is caught, and posting the gradient's or the fill's continuation is caught by the chained and floating rows. Then 331 of 331 at `-j16` and three of three under the 4-CPU cap.
+  - 13.19g stress past that: 2 CPUs at `-j32`, sixteen processes a core, failed six and eight suites, every one a time limit (waits of 5 to 20 s, QtTest's 300 s watchdog on `LargeCanvasBrushTests` and `LayerContextMenuEdgeTests`) but `SubjectRemovalFailureTests::aModelRunThatFindsNoMemoryIsARenderError`, whose 512 KB limit then found room; it passes alone at 2 CPUs three of three and under the 4-CPU load, and its cause under that stress is not established. `dev.sh test` and the distro checks run the suite one test at a time.
+  - 13.19g review: round 1 FAIL (the floating selection and the commands chained after a gradient still had a gap); round 2 PASS.
   - [ ] 13.19h The acceptance run and its shots.
   - [ ] 13.19i Distro checks: arch, fedora, fedora44, nixos, resolute.
   - [ ] 13.19j Tag `v1.4.5`; the release workflow publishes every package.