feat(diagnostics): observe WorkManager framework failures safely

[MrFr3di]

Closes #191

Purpose

Host-first, secret-safe WorkManager failure observation under #179. This PR remains draft and currently contains only O1.1: the platform-neutral typed failure contract plus its RED→GREEN evidence. WorkManager callback wiring, bounded recorder/store, Doctor projection and device acceptance are not part of the current implementation yet.

O1.1 TDD evidence

RED — observed

RED commit: 2516d8120c4c8535382c79654585b948b29daae2.

A real Gradle execution reached :core:common:compileTestKotlin and failed specifically because these production symbols did not exist yet:

• BackgroundWorkFailureKind;
• BackgroundWorkerCategory;
• BackgroundWorkFailureObservation.

The contract requires four stable WorkManager failure families, a closed worker-category vocabulary, no raw String/Throwable payload, deterministic value semantics and rejection of negative timestamps.

Minimal GREEN — verified

Implementation commit: 24b33931ce4ee09044e5ae467748de696d95a263.

Hosted-sync head: 7663cccb6d8eafeff4062be4a6ae4494e2056da3.

GitHub-hosted Hosted validation run 33048709750 completed SUCCESS on that exact head. O1.1 is therefore an accepted RED→GREEN checkpoint; the former self-hosted/no-verdict wording is obsolete.

Architecture boundary

The typed observation model remains platform-neutral and deliberately does not reuse generic diagnostic events capable of carrying arbitrary text. WorkManager remains owner of scheduling/retry/failure semantics. Later callbacks must be non-throwing and must never persist raw throwable text, URL/token/title or other secret-bearing data.

Remaining sequence

1. O1.2 RED for a bounded in-memory recorder/store;
2. minimal O1.2 GREEN: fixed capacity, newest-first snapshot, consecutive duplicate coalescing by typed identity with count/latest timestamp, no raw text/Throwable;
3. O1.3 RED/GREEN for the four Configuration.Builder failure callbacks behind a small adapter/factory;
4. coarse Doctor/export projection;
5. exact hosted validation;
6. canonical API26 + API36 smoke only when the Android adapter exists.

Boundaries

• no U0/test(tv): characterize rail/content geometry before UI fixes #189 production-source change;
• no chore(stack): stage Android stack refresh for runner validation #190 dependency staging;
• WorkManager stays 2.11.2;
• no retry/backoff/dispatcher/coroutine behavior change;
• no generic telemetry bus;
• GitHub-hosted CI only;
• canonical API26 + API36 only.

Do not mark ready or merge until the remaining O1 slices are independently RED→GREEN and the adapter has hosted/device acceptance.

[MrFr3di]

O1.1 exact-head GREEN checkpoint.

Current head: 24b33931ce4ee09044e5ae467748de696d95a263.

Self-hosted validation run 32997056529, job 98268962147 executed on the dedicated Windows runner and completed the substantive repository workload successfully:

• checkout / wrapper / preflight: PASS;
• dirty-runner admission + exact-head harness/device-routing contracts: PASS;
• two-AVD / repository truth / evidence-publication static contracts: PASS;
• Run MuxTV validation: PASS;
• runner reset/cleanup: PASS.

The workflow-level conclusion is red only because mandatory evidence artifact publication failed afterward under the separately owned storage-capacity problem (#209). Publication acceptance remains fail-closed; that transport/capacity failure does not invalidate the already completed O1.1 test/build workload.

This closes the executable O1.1 RED -> minimal implementation -> exact-head substantive GREEN loop. It authorizes the next TDD step (O1.2 bounded typed store RED), but does not make #207 ready to merge: O1.2/O1.3/Doctor integration and final acceptance remain pending.

Critical-path scheduling note: after this GREEN the singleton runner is being reserved for one exact-source U0 retry before another #207 push, so O1.2 will not be allowed to compete with U0.