Donner SVG Editor & Engine
SVG-native editor and embeddable SVG2 + CSS3 engine in C++20, with GPU (WebGPU) and compact CPU renderers, built for correctness, security, and performance.
Loading...
Searching...
No Matches
Codex handoff: preexisting drag/compositor defect cluster

For: a manual codex run (the deep compositor work kept killing Claude sub-agents with spurious "usage policy" API errors mid-task; route it through codex instead).

Goal: make bazel test //... fully green with zero disabled/skipped tests. Per the project's CLAUDE.md "Always-Green Main" policy there are no preexisting issues — every red test here is in scope, including a segfault. "Done" = full green via real production fixes, never by weakening a test.

Where to work

  • Worktree: /Users/jwm/Projects/donner/.claude/worktrees/fixdrag
  • Branch: v08/fix-drag-compose, HEAD 774b6ff1 (2 commits above base c0e8c53f).
  • Verify first:
    cd /Users/jwm/Projects/donner/.claude/worktrees/fixdrag
    git rev-parse --short HEAD # expect 774b6ff1
    git branch --show-current # expect v08/fix-drag-compose
    git status --porcelain # expect clean (ignore stray .ec* tooling files; rm them)
    If HEAD differs: git reset --hard 774b6ff1. Work only in this worktree — do not touch the main repo or sibling .claude/worktrees/* (parallel work edits PenTool.cc, EditorShell.cc, MenuBarPresenter.cc there).
  • Commit early and often on v08/fix-drag-compose; do not push. Re-check git rev-parse HEAD after each commit (a rogue background shell caused branch resets earlier; if HEAD jumps, reset to your last good commit).
  • Every bazel command must pass --disk_cache=/Users/jwm/.cache/donner-bazel-shared. Don't run concurrent bazel servers against this worktree.

The two commits already on the branch above base c0e8c53f:

  • 2087fbcb Fix translation-only drag compose-offset (preexisting compositor bug) — RC#1
  • 774b6ff1 Carry forward static-segment generation across unchanged hint-kind re-promote — RC#2

All edits so far are in donner/svg/compositor/CompositorController.cc.

What's already fixed (keep it)

RC#1 — immediate-layer re-raster clobbered the translation compose offset. A cheap promoted layer becomes an immediate (direct-render) layer. The rasterize-dirty-layers loop in CompositorController::renderFrameImpl (~line 2080) re-rasterized immediate layers on every frame because layer.isImmediate() bypassed the dirty check. On a pure-translation drag the fast path had already written the drag delta into canvasFromBitmap_ for bitmap reuse, but the unconditional re-raster called setBitmap, which resets canvasFromBitmap_ to identity and bumps the texture generation. Net symptom: layerComposeOffset() returned (0,0) and generation/rasterizeCount climbed 1→6 on a pure translation. Fix: skip the immediate re-raster when the layer is clean, payload-backed, has a stamp, and its canvasFromBitmap is a pure translation (immediateTranslationOnlyReuse). Also fixed hasSplitStaticLayers() (~line 1060): it early-returned true on the first ActiveDrag hint without counting total live hints, so ≥2 promoted layers looked like a single split target — now counts live (non-pending-demote) hints and reports split only when exactly one. Result: //donner/svg/compositor:compositor_tests is 50/50 GREEN (was red at base).

RC#2 — static-segment generation churn (partially addressed, VERIFY). rasterizeDirtyStaticSegments (~line 2694: staticSegmentGeneration_[i] = nextTileGeneration_++) and resyncSegmentsToLayerSet (~line 2790) minted fresh generations for reused/unchanged segment slots when a hint-kind change (Selection→ActiveDrag) re-promote shifted segment boundaries. Commit 774b6ff1 carries the generation forward across an unchanged re-promote. VERIFY it fully fixed compositor_golden_tests.SelectionToActiveDragDoesNotAdvanceUnchangedTileGenerations (was tile 5 generation 16 vs expected 12); finish it if still partial.

What's still RED (drive all to green)

Run targets individually first to reproduce, e.g.:

bazel test --disk_cache=/Users/jwm/.cache/donner-bazel-shared \
//donner/svg/compositor:compositor_golden_tests \
--test_arg=--gtest_filter='CompositorGoldenTest.SelectionToActiveDragDoesNotAdvanceUnchangedTileGenerations'
  1. //donner/svg/compositor:compositor_golden_tests
    • SplitBitmapsStableAcrossTranslationOnlyDragFrames
    • SelectionToActiveDragDoesNotAdvanceUnchangedTileGenerations (RC#2 — confirm fixed)
    • NonPromotedMutationInvalidatesOnlyContainingSegment
    • TightBoundedSegmentsPixelIdentityOnRealSplashWithDrag — RC#4: ~4190px mismatch. Tight-bounds segment compose recrops/shifts splash pixels under drag. Look at rasterizeDirtyStaticSegments / composeLayers applying the compose offset without recropping (the segment's source rect must move with the offset, not get re-cropped at the old bounds).
    • SplashLetterThenCloudOrbSelection
    • SplashCloudsDragMatchesReference
  2. //donner/editor/tests:async_renderer_tests — RC#3 + RC#5
    • ActiveDragStartDoesNotAdvanceUnchangedTileGenerations, SteadyActiveDragTargetReusesPublishedTextureMetadataOnly — RC#3: the editor GL-upload path in donner/editor/AsyncRenderer* / donner/editor/CompositedPresentation* re-uploads the drag-target texture every frame even when the tile generation is unchanged. It should publish metadata only (transform/offset) when the generation hasn't changed, not re-upload pixels.
    • RC#5 — SEGFAULT: this target segfaults mid-suite, and it segfaults at clean base c0e8c53f too (verified by swapping the base CompositorController.cc back in). So the segfault is independent of RC#1/#2 and must be root-caused on its own — run the target, get the crash, backtrace it (lldb, or a -c dbg build), fix it. This is in scope (no preexisting-issue exemption).
  3. //donner/editor/tests:editor_layer_stress_tests — writeback pixel diffs (3527–10401 px); likely downstream of RC#3/#4.
  4. //donner/editor/tests:rnr_replay_tests — FilterDisappearRepro3MatchesGoldenAfterSecondMouseUp, DeleteElementDoesNotResetPreviouslyMovedShapes.
  5. //donner/editor/tests:gl_rnr_replay_tests — GeodeDragZoomRerasterizesDonnerDOverlayEveryPresentedFrame.
  6. //tools/mcp-servers/editor-control:editor_control_session_tests — SplashOThenRDragKeepsStableSplitLayerPaintOrder (emits unknown:immediate:… translation_doc {0,0} tiles instead of layer:… tiles carrying the drag delta).

Likely remaining root causes, by area: RC#3 GL re-upload metadata-only (AsyncRenderer / CompositedPresentation), RC#4 tight-bounds segment compose recrop (CompositorController segment compose), RC#5 the async_renderer segfault. RC#2 may already be done. Several editor/ golden failures will likely clear once RC#3/#4 land.

Method (strict — CLAUDE.md)

  1. Reproduce each failure first (run the target, capture the assertion or the crash backtrace).
  2. Root-cause it in production code; the tests encode the correct contract.
  3. Confirm red→green for that target; commit.
  4. Never edit/disable/skip/delete a test to make it pass.
  5. "Anti-aliasing" is a banned root-cause for any pixel diff — the pixelmatch harness already excludes AA, so any flagged diff is a real bug (wrong transform / offset / crop / generation / premultiply). A 4190px mismatch is definitely not AA.
  6. Remove any debug fprintf/scratch before committing. Run /opt/homebrew/opt/llvm@18/bin/clang-format -i on edited files. Respect the destFromSource transform-naming convention (e.g. bitmapEntityFromEntity, canvasFromBitmap).

Final gate

bazel test --disk_cache=/Users/jwm/.cache/donner-bazel-shared --flaky_test_attempts=2 //...

→ 0 failures. If you can't reach full green, commit all real progress and report exactly which targets remain red with assertion/backtrace evidence and a next-step hypothesis — do not fake green or relabel anything "preexisting/out of scope".

When done

Report the final v08/fix-drag-compose HEAD SHA, each target turned green with its root cause, the segfault root cause, and the full bazel test //... summary line. Then this branch merges into v0_8_drive (squash-and-merge per CLAUDE.md when it eventually lands on main).