fix(ci): stabilize responsive overlay interactions

[nocoo]

Current verification (2026-10-01): ordinary CI 36796771419 passes all 261 cases on the first attempt. Stress 36796811662 fails all five cells on the normal-motion WebKit dialog-close assertion; cell 5 also has a card-visibility flake before interaction. All newly added touch and reduced-motion regressions pass first-attempt in all six reports. This Draft remains blocked; no merge or post-merge verification has occurred.

Outside touch can close a card menu, restore its trigger focus, then generate a compatibility mousedown that removes that focus. The supported outside-pointer callback now cancels the original primary touch PointerEvent default while allowing normal immediate Radix dismissal. It does not cancel the custom dismissal event or depend on a later click. Mouse, keyboard and non-primary behavior are preserved.

Creation-dialog content also honors reduced motion in both open and closed states. The existing app override now keeps class-plus-attribute specificity against the normal entrance rule and overrides the higher-specificity closed-state utility. A regression using the actual built CSS fails on the old closed-state animation and passes with the correction, while normal entrance/exit animation and real Escape/focus behavior remain verified. No normal animation is disabled.

The branch also isolates globally unique fixture slugs, retains complete per-width keyboard checks and same-page resize continuity, and separates runtime-motion behavior from an overloaded compound case. Overlay failure attachments distinguish assertion-catch from after-test-function because automatic failure screenshots precede afterEach. All original timeouts, retries, four CI workers, motion behavior, exact selection assertion and flaky-test gates remain unchanged.

Validation on this correction: a native held-touch regression reproduces lost trigger focus after touchEnd with original application behavior; the corrected light/dark cases pass. Existing Chromium and macOS WebKit card interactions pass 50/50 on their first attempt. Normal hooks pass 3,492 unit tests, 197 in-process integration tests, types, lint and secret scan. All 259 prior candidate cases remain plus two browser reduced-motion CSS regressions, total 261. Normal push runs 222 real HTTP tests and root dependency scanning; separate root/Worker/CLI scans report no issues.

Source evidence: https://github.com/nocoo/zhe/actions/runs/36357254553 (stable main Stress), https://github.com/nocoo/zhe/actions/runs/36790306239 (native touch focus race), https://github.com/nocoo/zhe/actions/runs/36793727979 (superseded click-wait regression), https://github.com/nocoo/zhe/actions/runs/36795187899 (all card/touch cases first-pass; separate normal-motion pending exit and reduced-motion cascade failure). All commits are retained; the superseded listener state machine is removed in the current tree.

Pointer cancellation suppresses compatibility mouse events, not click/auxclick/contextmenu. The native regression verifies actual focus behavior; this is not a general guarantee against background activation. Current-head ordinary CI and all five Stress reports must show all 261 cases passing on the first attempt before merge, followed by fresh main verification. The separate hosted WebKit normal-motion pending-exit issue remains unproven, and the PR stays Draft while gates are unresolved. No workflow, secret or environment protection is changed.