[mcdonc]

Summary

Fixes the recurring tap_widget failed on s_9: web_gesture_not_supported flakes in the scheduled main E2E: fmtk Frontend Tests (workspaces) runs — 1 test on Oct 25 (run 40182792728), 5 tests on Oct 18 (run 39116591202). Closes #3588.

All the failures tap the shared-workspace card via tap_button_exact(name) after switching to the Shared with Me segment. The lookup picked the first button-typed node carrying the name as a whole line — and that node did not expose SemanticsAction.tap, so fmtk's tap_widget refused the gesture. A control's semantics can settle label-first (the button flag and label land before the tap action registers), and a non-tappable button-typed node can precede the real one in tree order; in either state hits[0] was never a usable target, because tap_widget can only tap nodes exposing the action.

tap_button_exact now prefers hits that expose the tap action and waits out the settle window (10s) before failing, with an error that names which of the two states it saw. All 42 call sites across the suites get the hardening.

Verification

• test-fmtk-e2e src/frontend/e2e-tests/fmtk/test_terminals.py (fresh stack): 5/5 pass, including both Oct-18 cascade tests (test_gate_own_terminals_by_role, test_shared_terminal_spectator_watches).
• test-fmtk-e2e src/frontend/e2e-tests/fmtk/test_sharing.py: the issue's failing test test_acl_editor_add_save_and_persisted_effect passes. (The module's chained scenarios 4-6 are locally load-sensitive when run standalone — see its docstring; CI runs the authoritative sequence.)
• test-push: green.
• The fmtk-e2e-workspaces workflow dispatched on this branch: see checks below.

[mcdonc]

Fresh-eyes review

Blocking issues

None.

Verification of prior findings

1. Stale "shared tiles surface no tap actions" docstring — fixed. test_terminals.py:107-113 now reads "its merged semantics can surface the label before the tap action registers … when that happens (Fix fmtk E2E (workspaces) flake: tap on non-tappable merged button node (run 40182792728) #3588) — tap_button_exact waits out that window." The false steady-state claim is gone from the entire tree (rg "no tap actions|surface no tap" → zero hits), and the replacement documents the fix's actual behavior instead of contradicting it.
2. Mechanism stated as fact — fixed. fmtkharness.py:2204-2215 now gives both hypotheses ("consistent with … semantics settle … or with a non-tappable labeled node preceding the real one in tree order; the nightlies' artifacts do not distinguish the two") and grounds the code's agnosticism ("Both states are unusable targets"). This is exactly the hedge requested.
3. Accepted nits — all applied and correct:
   • FmtkError-is-load-bearing note at fmtkharness.py:2215-2217; accurate — the sole catching call site is test_sharing.py:266-267 (verified every other except FmtkError in the suite wraps wait_for_*/dart_logout, not tap_button_exact).
   • timeout=2 at test_sharing.py:266: failure inside the catch branch now bounded ~2-3s (one snapshot + sleep + snapshot past a 2s deadline) instead of 10s. No correctness loss — a raise there is caught, the outer loop re-reads state (has_text("Popup menu") / Add-ACE check) and retries; the success path is unaffected (timeout only bounds failure).
   • Hit-scan extracted to button_hits (fmtkharness.py:2234-2240); comprehension body identical to the inlined original — behavior-preserving. tap_button_exact drops back to xenon rank A (5 branches), button_hits is 3.

Regression bleed

None. git diff --stat 41e54bb0e..HEAD confirms the delta touches only the three reviewed hunks; the loop's selection, raise messages, and both FmtkError shapes are unchanged. button_hits has no name collision (2 grep hits: def + sole caller). Contract suite re-run on HEAD: scripts/tests/test_fmtk_harness.py — 23 passed.

Nits / questions

• test_terminals.py:109-111 still tells only the same-node-settle story ("label before the tap action registers") while the harness docstring carries both hypotheses. It's hedged with "can" and cites Fix fmtk E2E (workspaces) flake: tap on non-tappable merged button node (run 40182792728) #3588, and the climb-failure it describes is real under that story — acceptable, but a "or a non-tappable twin precedes the button" clause would make the two docstrings fully consistent.
• button_hits is now public API on FmtkClient with one caller; fine under the repo's no-leading-underscore rule, just noting the surface grew by one method.

Verdict

APPROVE — both important findings are genuinely fixed (stale claim eradicated, mechanism hedged, catch-dependency documented), the accepted nits are correctly applied, and the refactor is behavior-preserving with the contract suite green on HEAD.

[mcdonc]

3D dependency graph (self-contained HTML)

• Render in browser
• Gist source

The page is fully self-contained: save it from either link and it opens offline. GitHub serves the raw gist as plain text, so the source link downloads; the render link views.

[mcdonc]

Successfully created backport PR for stable/2.0:

• Fix fmtk E2E workspaces flake: tap_button_exact waits for a tappable node (#3588) #3590