macOS gesture regression rejects the visible slider center before provider dispatch

[RolandUI]

Full CI 36072375163 on b90ca39 fails macOS Runtime job 107882571765 in BridgeHeadlessSmokeTests.SemanticGesturesUseRangeProviderAndBoundsDerivedPointerFallback, line 4399. The first provider-backed drag of gesture-slider returns failure: "Gesture source center is obscured or not hit-testable." The test fails after 19 ms; this is not evidence of an IPC deadline. The immediately preceding depth-8 tree query succeeded and supplied the slider node. Aggregate: 816 passed/1 failed/5 explicit native skips (822 total, 8m10s).

Retained evidence: artifacts/agent-qa/expanded-campaign/ci-b90-macos-failed.log. Windows full tests and all three native QA labs passed; Linux independently failed #184. #179 expression and #181 matrix tests are not the failing macOS cases in this run.

Acceptance:

• Retain the first hosted failure and capture useful target bounds, layout/render readiness, center hit-test and rejection diagnostics in the affected fixture without weakening input guards.
• Reproduce or identify the headless/macOS layout/compositor or test-isolation cause; distinguish a legitimate obscuration refusal from stale readiness assumptions.
• Provider drag still updates the range once; fallback paths, threshold slider, nested target and negative/obscured input guards remain covered.
• Focused local/macOS validation and the affected full CI test gate pass; report unavailable reproductions explicitly.

Queued behind active #183. #172/#173/#174 fixture expansion/campaign remains unfinished. No release/publication.

[RolandUI]

The same defect reproduces again in full macOS Runtime job 107954766598 of before-fix CI 36096369354 / 8bebfe6. SemanticGesturesUseRangeProviderAndBoundsDerivedPointerFallback rejects the first provider slider at line 4399 with the same obscured/not-hit-testable message, now after 12 ms. The #188 InputHitTest correction therefore did not resolve this readiness case. Full aggregate: 854 passed/two failed/five skipped, 9m56s; the separate second failure is #184's controlled queued Unix connection reproduction. Raw/clean ci-8beb-macos logs are retained under artifacts/agent-qa/expanded-campaign. All three native QA jobs passed with observed scale 1. #184 remains the sole active issue; #185 follows its local fix validation. No release.

[RolandUI]

Starting as the sole implementation issue after #184 fix 8bf3c4b moved to review. The original and fresh hosted macOS failures are retained (19 ms and 12 ms, first slider drag). Inspect actual frame readiness, layout/bounds and center hit path before the provider request; use Avalonia 12.1.3 public headless rendering APIs to establish a real frame precondition if confirmed. Preserve hit-test/obscuration guards and all provider/fallback/capture/negative cases. Validate affected local cases and the full hosted macOS gate together with the #184 Unix after-fix regression. Do not claim the original uninstrumented compositor sequence is known. No release.

[RolandUI]

Pushed 3b82d1d: the affected headless fixture records initial frame/layout/center-hit diagnostics, explicitly captures/requires a real 500x430 frame and slider hit before the first input, and verifies exactly one range change. Avalonia 12.1.3 public compositor/headless sources show why draining dispatcher jobs is insufficient as a frame precondition. Production input guards, deadlines and fallback/capture/negative cases are unchanged. Linux/macOS CI now retain per-test TRX diagnostics.

Validation: instrumented cold local before run passes in 9 s with a frame already available, so it is not a local reproduction. After correction all 36 affected gesture/explicit-input/picking/action-explanation cases pass (one minute, clean build), with expected frame, bounds and hit diagnostics in gesture-readiness-after-results/gesture-readiness-after.trx. The original/fresh macOS 19 ms/12 ms failures remain retained; their uninstrumented compositor sequence is not invented.

Moving to review, not closing. Full combined CI 36100032601 runs exact 3b82d1d including #184's Unix fix: https://github.com/RolandUI/AvaScope/actions/runs/36100032601 . Do not cancel for subsequent capability work. Continue #174 while hosted validation runs. No release.

[RolandUI]

Hosted checkpoint on 483fedb, CI 36102961992: the complete run is failed only on new #193's macOS action fixture readiness, before its test body. Windows passes 856 plus six explicit skips; Linux passes all 33; macOS has 856 passed/one failed/five skips. Downloaded TRX verifies #184 queued Unix connection retention on Linux/macOS (269/299 ms), #185 macOS gesture readiness/hit evidence (1.583 s), both Unicode variants and the controlled cooperative-budget regression on Windows/macOS. Windows's eight routed input spans are 0.306-20.449 ms with all expected Unicode values and exactly four input/change events per variant. All six native Direct/standalone lab runs pass, with clean 483fedb identity and owned termination verified; all scale 1x. macOS downstream package/workflow stages were skipped after the single failure.

Evidence is retained under artifacts/agent-qa/expanded-campaign/ci-483-{windows,linux,macos,native}; compact report is committed at 2f01918. These results validate the named cases on independent hosts, but do not establish #191's original slowdown cause or repair #192's missing lifecycle phase evidence. #193 is now the sole active defect; #174 is ready for continued coverage after that slice. No release.