test(e2e): capture scenario assertion failure evidence

[billsbooth]

Summary

Capture failure artifacts when a live scenario assertion fails outside a CLI step. The iOS long-press search exhausted allowed probes, then threw without preserving the failed screen or accessibility tree.

Scope

Two files: shared live-scenario harness and its deterministic tests. Capture before caller teardown, preserve the original exception, avoid duplicate capture for already-reported command failures, and leave successful scenarios unchanged. No gesture, driver, timeout, recovery or product behavior change.

Validation

At 6ef99b084, pnpm check:affected --run passes all runnable gates: format, lint, types, Fallow and 101 Node integration cases (9 native skips). Six new cases cover allowed misses, pre-command assertions, duplicate suppression, unavailable capture, artifact I/O failure and success; three failed before implementation. CI pending. This does not fix the original iOS visibility failure.

Verification

iOS failure contains four unsuccessful visibility probes and cleanup, but its uploaded fixture artifacts lack failure screenshot/snapshot. Source confirms only unexpected command failures previously invoked collection. The new tests assert original-error identity, owned-session flags, artifact output and coverage bookkeeping.

Automated Test Plan

Run node scripts/node-test-tmpdir.ts --test test/integration/live-device-e2e-runtime.test.ts test/integration/live-device-e2e-failed-step-evidence.test.ts. All ten tests must pass with no device or account. Automatic native fixture lanes remain selected: on any scenario assertion, inspect failure report and capture artifacts before session cleanup; successful runs must create no failure captures. No production data, funds or external delivery required.

[thymikee]

The change looks good at 6ef99b0: scenario assertions now reuse failure capture before cleanup, preserve the original error, and avoid capturing command failures twice. The deterministic regressions cover those boundaries. CI remains the blocker: iOS failed in the alert XCTest before fixture E2E, and Android failed at automation-press. Neither run verifies the new scenario-assertion capture path; confirm the uploaded screenshot and snapshot on a deliberate assertion failure when that lane can run.

[thymikee]

I refreshed this change on current main in #2921. The gap remains: scenario assertions after allowed CLI misses still lost their screenshot and snapshot. The replacement keeps the original implementation commit authored by @billsbooth, shares the command/scenario failure report formatter, and covers the newer platform-owned device evidence hook.

The two-file replacement passed 11 focused tests and pnpm check:affected --run. A deliberate local iOS 26.2 simulator assertion after an allowed miss preserved the original error and captured a valid PNG and Settings snapshot before cleanup. New CI is pending on #2921.

I tried updating this fork branch with an exact-head lease, but GitHub rejected the rebased push because the maintainer's OAuth token lacks workflow scope for workflow commits now on main. #2921 carries the same work without changing the contributor's fork, so I am closing this PR as superseded.