Preserve Firebase polling diagnostics and collect terminal iOS results

[leehack]

Problem

The PR #515 remote controller loses the actionable polling failure and collects results before cancellation has reached a terminal state. A bounded iOS attempt therefore ended with no test execution, collection=COMPLETE, and only a generic StateError diagnostic. Qualification correctly stayed false and cleanup succeeded.

Reproduction evidence

• Runner commit: 21f89f94d is the current reviewed head; the submitted run began on b92a0e1df. Both retain the same controller code.
• Clean signed bundle: 31867c10c, native v0.4.1, Qwen3.5 GGUF Metal.
• Target: Firebase iPhone 16 Pro / iOS 18.3, Xcode 26.2, eight-minute test timeout, no flaky retries.
• Matrix: matrix-3k4vi693lyhyc; run: qa-1789730285358167.
• Journal: FAILED, cleanup VERIFIED, collection COMPLETE. Error: Remote operation failed (StateError); inspect bounded local diagnostics.
• Collected matrix.json is still PENDING; only the uploaded tests.zip was available at collection time.
• Fresh provider read: matrix FINISHED / outcome INCONCLUSIVE; execution CANCELLED. Tool Results: complete, abortedByUser=true, no testProcessDuration, no XCTest output. The controller's cleanup issued cancellation; this is not evidence of an iOS runtime/model failure.
• Android's earlier run completed collection normally but failed its model download deadline: 2,285,638,612 / 2,588,147,712 bytes in ten minutes, before inference. That is a separate model-delivery gap.

Expected follow-up

• Persist safe structured polling diagnostics (phase, operation, HTTP status or typed failure code) without tokens, URLs with credentials, or arbitrary secret-bearing error text.
• Exercise transient polling failures explicitly; any bounded retry must retain the existing absolute deadline and must never resubmit the matrix.
• Refresh the matrix/result inventory after terminal cleanup and distinguish a transferred pre-terminal snapshot from complete final test artifacts.
• Keep cancelled/preparation-only runs unqualified. Preserve verified cleanup and fail-closed allowance handling when test-process usage is absent.
• Add production controller-path positive and negative tests before another bounded iOS attempt.

No automatic rerun is requested. Both matrices are terminal and no GCE VM was created.

[leehack]

Implemented in #515 at 89097b704fbfe90718484cd577076e7a5cb7083d, with manual recovery follow-up 3d59bb2dbd4ad904fa7147d8f94bfb3cccf6c025:

• Safe typed authentication/HTTP/timeout/transport/JSON diagnostics; no provider bodies, tokens or credential URLs.
• At most three consecutive read-only polling attempts under the original absolute deadline; no submission retry.
• Failure cleanup precedes terminal Firebase artifact collection; preterminal snapshots cannot be COMPLETE.
• Manual cleanup invalidates old collections and refreshes status. Manual collect refreshes terminal status. Failed recovery persists incomplete/unqualified journal and summary.
• 42 provider regression tests pass (one existing x64-only ABI test skipped on Mac); independent review covers cancelled/preparation-only state, missing assertions and failed recollection.

Live verification: recovered original matrix matrix-3k4vi693lyhyc to actual FINISHED / INCONCLUSIVE / CANCELLED (unqualified). One bounded current-source tiny iOS retry on iPhone16Pro / iOS18.3 completed: matrix matrix-2j9mb3xdi1y7a, actual test duration 43 seconds, provider SUCCESS, assertions passed, collection COMPLETE, cleanup VERIFIED, qualified=true. This is public-API tiny-model CPU infrastructure evidence, not Gemma/Qwen/Metal qualification. No duplicate submission or paid retry. Fresh usage inventory records one rounded minute for the retry; previous missing-duration cancelled execution is still conservatively reserved, not treated as free zero usage.

Firebase result. Keep this issue open until the PR is merged; latest recovery follow-up CI is running.