fix(android): report dispatched no for adb host refusals

[thymikee]

Summary

An adb failure that adb itself prints before any device-side work now reports dispatched: "no". The classifier marks a host refusal (device_unauthorized, device_offline, multiple_devices, no_devices, device_not_found) only when the whole adb output is the refusal: empty stdout, a first stderr line that is exactly the refusal, and only adb's fixed trailer lines after it. Mixed output, server_version_mismatch, install_*, connection_dropped and timeouts stay unknown. attachAdbFailureHint stamps details.adbHostRefusal. discloseAdbInputDispatch and the one-shot touch helper read that stamp, so there is no string matching at the producers.

Closes #3075. Part of #3069. Stacked on #3071. Touches 6 files.

Validation

Tested commit: df909a1. Not device-facing, so no live evidence.

Every lane passed except vitest-related. There, 5 of 1324 test files failed, each with a 5000ms timeout, and each passed when rerun alone. The passing lanes were format, lint, typecheck, layering, di-seams, fallow, mcp-metadata, build, package, integration-node and macos-coverage. The 5 files are:

• runner-recovery-wiring
• runner-session-lifecycle
• runner-session
• apple-ci-impact
• screenshot-density

screenshot-density passed alone, but took 4.95s against the 5s timeout. The same file is untouched by the branch diff against 6165953, so this is a slow test under host load rather than a regression. No code changes or commits were made.

Contention reruns passed alone:

• packages/platform-apple/src/runner/tests/runner-recovery-wiring.test.ts
• packages/platform-apple/src/runner/tests/runner-session-lifecycle.test.ts
• packages/platform-apple/src/runner/tests/runner-session.test.ts
• scripts/tests/apple-ci-impact.test.ts
• packages/platform-apple/src/core/tests/screenshot-density.test.ts

Unresolved risk: fill at zero dispatched steps (text-input.ts) is unchanged. The one-shot helper failure now also carries the classifier's adbFailure, hint and retriable.

[cubic-dev-ai]

All reported issues were addressed across 96 files

<sub>Reply with feedback, questions, or to request a fix.

Fix all with cubic | Re-trigger cubic</sub>

[thymikee]

[claude-fable-5-1] responding on behalf of @thymikee

All three findings addressed in 1c7f3cd.

1. Mixed-refusal fixture row (P2) -> 1c7f3cd. The driver now feeds error: device offline\nKilled through the real tap executor and asserts unknown; trigger text updated. Mutation: trailer.every(...) -> true in classifyHostRefusal makes the row report no.
2. touch-helper parsed-final/nonzero branch (P2) -> 1c7f3cd. The error now goes through attachAdbFailureHint before disclosure (still unknown). New test: nonzero exit with adb: device offline keeps adbFailure, hint, retriable. Mutation: removing attachAdbFailureHint from that branch fails the test (verified).
3. adb-failure.test hint (P3) -> 1c7f3cd. The expected hint is a hard-coded literal. Mutation: changing the hint text in adb-failure.ts fails the assertion.

[github-actions]

Size Report

Metric	Base	Current	Diff
Installed (including dependencies)	4.91 MB	4.91 MB	+919 B
Package (unpacked)	4.91 MB	4.91 MB	+919 B
Package (download)	1.47 MB	1.47 MB	+382 B

Startup median (7 runs, lower is better):

Scenario	Base	Current	Diff
CLI --version	21.6 ms	23.1 ms	+1.5 ms
CLI --help	63.2 ms	66.8 ms	+3.6 ms

[thymikee]

This PR is ready at 1c7f3cd. The code looks correct to me, and all 19 checks pass. The PR body reports five Apple/screenshot vitest timeouts under host load. Those files do not overlap the Android adb-failure and touch-helper paths this diff touches. There are no conflicts.

Not blocking, and you can take or leave these: (1) HOST_REFUSAL_SHAPES re-spells the five patterns that ANDROID_ADB_FAILURE_MATCHERS already holds. If someone edits a matcher regex, the anchored shape can fall behind. The result stays unknown, which is safe. Declaring the anchored head pattern on the matcher entry would leave one table. (2) HOST_REFUSAL_STDERR in dispatch-disclosure.test.ts and HOST_REFUSALS in adb-failure.test.ts are near-identical. One exported fixture table would serve both. (3) The PR body could say that server_version_mismatch is excluded because adb retries the command after it restarts the server. The "Unresolved risk" line about fill at zero dispatched steps looks moot, since fillAndroid counts the focus tap before any IME broadcast.

I did not run the tests or the mutation (trailer.every -> true). I relied on green CI and the diff. I did not check against a real unauthorized or offline device, so I cannot confirm that adb emits exactly the asserted strings. Different wording falls to unknown, which is the safe side. I could not prove one race from code. A device that drops after the shell stream opens would have to print a bare error: device offline or device not found with empty stdout to be misclassified. The adb client only prints those at transport selection, before the command is sent. I could not confirm this for remote or cloud ADB providers that relay stderr. I also did not check readers of adbFailure or retriable outside packages/platform-android.

[thymikee]

[claude-fable-5-1] responding on behalf of @thymikee

Replayed onto #3071's d7b442d040; the rebase applied cleanly with no conflicts. Added refactor(android): one adb failure table carries the host-refusal shapes: HOST_REFUSAL_SHAPES and ANDROID_ADB_FAILURES_BY_REASON are folded into ANDROID_ADB_FAILURE_MATCHERS, where a row may carry an anchored hostRefusal pattern and the classifier reads it from the row it matched. The existing adb-failure, dispatch-disclosure and adb-provider-scope tests pass unchanged (none imported the removed constants).

[thymikee]

The latest changes at cf00361 look good. The rebase left only one logical change, the table fold in the adb failure classifier, and it matches the earlier review of 1c7f3cd. The existing adb-failure and dispatch-disclosure tests cover that route. I did not run the platform-android tests locally, and I did not run a live adb device. I compared the two table versions side by side. I also did not re-review the base changes from #3071 that came in through the rebase. All 19 checks pass. There are no conflicts. Nothing more is needed from review. Merge can go ahead once #3071 lands.