fix(ci): retry API 37 infrastructure crashes

[attmous]

Summary

• add a fail-closed managed-device runner that retries API 37 only for zero-test instrumentation startup crashes or the exact Android Bluetooth hardware-error abort
• retain every failed infrastructure attempt and stop after two retries; assertions, identified MusFit process crashes, and exhaustion remain red
• apply the same policy to ordinary and coverage journeys and expand only their hard timeout ceilings

Root cause

PR #178 fixed the stale localized assertions: hosted API 28 passed all 10 journeys and the repaired API 37 assertions passed. The exact merge-SHA device job then failed twice in the Android 37 emulator itself:

1. Android Bluetooth aborted with Hardware Error Event with code 0x42 after two passing tests.
2. A manual rerun produced a valid zero-test report with Instrumentation run failed due to Process crashed before any test began.

Verification

• ./scripts/dev/test-dev-workflow.ps1 -SelfTest
• ./scripts/dev/verify-musfit.ps1 -Preset Full -RetryOnGeneratedOutputIssue — 2,158 tasks passed; isolated CycloneDX SBOM passed
• final wrapper command on musFitApi37InternalDebugAndroidTest — 10 tests, 0 failures/errors/skips
• workflow YAML parsed locally and git diff --check passed

Risk and rollback

Risk is limited to the post-merge device lane. Product/test failures still fail immediately, infrastructure failures remain bounded, and all failed attempts are uploaded. Rollback is the single commit 8e228402473948c67d4db87163dce6df2aa7ad6f.