In-process Android test driver, pluggable drivers, unified tracing, and multi-run report views

[handstandsam]

Summary

Syncs internal changes through 2026.09.01. The headline items:

• An in-process Android test driver (in development) (trailblaze-android-test) that runs Trailblaze tools inside the app's own AndroidJUnitRunner test — Espresso for Views, the app's existing Compose rule for composables, with two native selector dialects (androidView, compose) over one non-overlapping hybrid tree. Reachable from the host CLI, and runnable against an app with no Gradle module of your own via trailblaze inprocess make-test-apk.
• Drivers are now plug-ins. HostDriverDescriptor replaces when (driverType) arms that were scattered across device discovery, the YAML runner, screen capture and three device listings. Adding a driver is three edits; an app declares which drivers it has rather than spelling each one out everywhere. Compose desktop, both Playwright web drivers and Revyl are converted.
• One tracing pipeline. Four overlapping profiling systems collapse into one span producer, one durable timing record, one consumer (trailblaze profile), and two export formats — Chrome Trace always, OTLP via trailblaze otel on demand. Host and device halves merge into a single trace.json on one trace id.
• Multi-run report views. Several runs of the same trail render as one report with Map, Grid and Replay projections — every lane advancing on a shared wall clock, playing each lane's video where one was recorded. Plus run comparison, snapshot baselines against a previous run, and attachment rendering across every report surface.
• trailblaze usages --json — a documented, versioned report contract for gating CI on which trails a tool change affects.
• Session video is now opt-in (--capture-video / TRAILBLAZE_CAPTURE_VIDEO / capture-video), and the legacy Compose/WebAssembly report is removed.

Also: multi-device runs take per-run --bind / --configuration flags instead of daemon-wide env vars, Playwright web recordings replay trustworthily and the module's tests run in CI, and device-clock skew no longer corrupts a run's recording.