Pointer diagnostics reports the window fallback border instead of the actual hit target

[RolandUI]

Actual native #172 exploration on 8e36e2a exposed a reproducible pointer diagnostic failure on the expanded shared fixture. Windows / Avalonia 12.1.3 / standalone provider 47034aa / scale 1 / client 1120x800. The original failed evidence is retained.

Reproduction:

1. Launch the standalone native QA app, select Input (tab 6), find qa-pointer-pad.
2. Its observed bounds are 44,315,520,150 in top-level DIP. Explicit synthetic click and CLI native drag succeed; independent journal/native pixels show two presses/two releases/one drag at delta 80,30, with no held input.
3. MCP pointer_diagnostics with move/assert_hit at center 304,390 rejects qa-pointer-pad.
4. A fresh actual CLI pointer-diagnostics at 150,390 (clear pad area, away from its non-hit-testable text) fails identically. The bridge move metadata reports the expected pad node visual:2984e99 at bounds 44,315,520,150, but the diagnostics path ends at PART_TransparencyFallback (visual:184a24a, full window bounds), excluding the actual pad. CLI exit 1, structured error runtime_pointer_input_hit_path_mismatch, followed by runtime_pointer_expected_node_not_hit.

Evidence: artifacts/agent-qa/keyed-standalone-01/qa-run.json, artifact-identity.json, D001-D005 positive/input-guard results, D004-drag-state.json, D006-pointer-diagnostics.arguments.json / .result.json, D007-cli-pointer.request.json / .stdout.json / .stderr.log / .exit.txt. Native pad/Unicode/focus images and independent journal are retained. Session c4b0dd0dc8cf41c9adb64aaade7567f8, owned PID 20960; stop state will be recorded separately.

Likely area: RuntimePointerDiagnosticsRunner reconstructs a path from rectangular tree bounds rather than respecting Avalonia's actual hit path. Confirm ordering/visibility/clipping/transform and depth-limit behavior before choosing the fix; do not special-case the QA window or silently discard mismatches. The second reproduction rules out the center TextBlock as the sole cause. This is distinct from #185's macOS gesture guard and must not be conflated with it.

Acceptance:

• Reproduce the incorrect path with meaningful real-Avalonia and/or controlled tree regression coverage.
• Prefer current public Avalonia hit-test evidence where available; preserve explicit unavailable/partial/unsupported coverage instead of reporting guessed bounds as a verified hit.
• Preserve true overlay/occlusion, popup ownership, clipping/transforms and mismatch diagnostics; no QA-specific control-name workaround.
• Actual native CLI/MCP positive and negative assertions agree with independent input counters/pixels on both fixture integrations.
• Relevant focused/full validation and issue/commit handoff pass. No release.

[RolandUI]

Read-only source confirmation: RuntimePointerDiagnosticsRunner.AddHitPath sorts containing child rectangles by ascending area and chooses the first; it does not use Avalonia hit-test ordering or visibility. In both retained failures the chosen full-window border is PART_TransparencyFallback. This explains the reconstructed path, while the separate CLI move metadata at 150,390 correctly identifies qa-pointer-pad. No fix yet. The standalone session was reset successfully and its owned PID 20960 terminated; stop.stdout.log and final-reset-state.json are retained. Direct fixture exploration follows before selecting the defect for implementation.

[RolandUI]

Starting #186 as the sole in-progress issue after completing #187 (91819d1).

I will replace rectangle-order hit-path guesses with the existing generation/geometry-pinned public pick evidence, retaining explicit unavailable/partial coverage and true mismatch diagnostics. Primary-window DIP coordinates must be mapped correctly when considering another registered top-level; popup ownership and unverified native overlap remain visible.

Validation: reproduce the fallback-border error using actual Avalonia controls; cover overlays, clipping/transforms, incomplete paths and unavailable picking, plus existing controlled Core/CLI pointer cases. Then rerun native CLI/MCP positive/negative journeys against independent counters and reviewed pixels on both fixture integrations. Full applicable validation will follow before closing; no release.

[RolandUI]

#186 checkpoint (uncommitted implementation; sole in-progress issue): the real-Avalonia transformed-pad regression fails before the fix and the 17 affected picking/Core/CLI/protocol cases now all pass (pointer-after.trx, 36s). Full Release build is clean. Actual paths come from generation/geometry-pinned public picking; nearest rectangles are separately labeled, incomplete/ambiguous assertions refuse, and cross-root/rendered-pixel mapping is covered.

Native Direct pointer-direct-01 reports the correct pad path. Clear-pad CLI/MCP positives, shallow-tree coverage, wrong-node/outside-pad negatives, native click/80x30 drag counters and stale zero-event refusal pass. Reviewed native/public images agree. Center movement now exposes a genuine legacy input mismatch, filed separately as #188; it remains in diagnostics. Original false fallback-border paths are still retained.

A real popup was observed but disappeared before its input step; this is explicitly incomplete. Agent-authored wrong revision/unsupported menu requests were safely refused and corrected. Direct was reset/stopped (owned PID12868), persistent client2964 exited0 after23 calls. Full local Release suite remains active; standalone and isolated native popup validation follow. No completion claim or release.

[RolandUI]

Implemented in 36cc427. Pointer diagnostics now uses the existing public generation/geometry-pinned pick API, preserves root-to-leaf paths, reports unavailable/partial coverage, maps cross-root desktop coordinates, refuses ambiguous/incomplete assertions, and places markers in the selected root's rendered pixels. Nearest-node rectangle distance remains separate. True legacy input mismatch #188 stays visible.

Validation:

• Real Avalonia transformed/overlaid/clipped regression fails with the old path; all 19 final picking/Core/CLI/protocol tests pass (32 s), including mapping loss and primary-window movement.
• Full Release build: zero warnings/errors. Full Release suite: 838 passed, one failed (Full Release launch timeout test exceeds its child-lifetime timing guard #189 launch timing guard), five native skips, 35m05s. The two final mapping tests were added after this full run started and passed in the separate focused run. Combined gate is not green.
• Native Direct pointer-direct-01 verifies actual pad paths, clear-area CLI/MCP positives, negatives, native click/drag, zero-event stale refusal and reviewed pixels. Center exposes independent Legacy synthetic pointer input targets hit-test-invisible visuals #188; it is not hidden.
• Fresh isolated native pointer-standalone-01: 18 retained checks; actual click/CLI drag counters 2/2/1, delta 80x30, exact popup point 81,20, mapped owner point 125,244, ambiguity/owner-only refusals, no move-only journal mutation.
• Fresh isolated native pointer-direct-02: eight retained popup checks, CLI local assertion/marker, MCP mapped ambiguity/owner-only refusals, independent journal and reviewed Windows popup pixels.
• Both fresh apps reset and owned termination succeeds (11868,11992); clients exit 0 (4444:19 calls;21184:13 calls), empty stderr. All observed native scales are 1; no Retina claim. Original interrupted-popup attempt remains retained without an inferred cause.

Moving to review pending the fresh combined gate after #188/#189 and the queued platform failures are addressed. See docs/AGENT_QA_PHASE_ONE_REPORT.md. No release or version change.