[BE-0283] feat(android): add network-capture assertions to the adb backend

[0x0c]

Summary

Ports request / requestSequence network-capture assertions to the Android (adb) backend, closing a
gap since BE-0007: the runner already pre-starts a NetworkCollector per Android device and injects
BAJUTSU_COLLECTOR into the launch env, but nothing on Android ever read it, so a request assertion
silently degraded to "zero exchanges observed" instead of failing loudly. This merges BE-0283 to
Implemented.

BajutsuAndroid gains an OkHttp interceptor (BajutsuNet.kt) that reports each exchange to the same
in-process collector iOS uses, and the runner bridges the emulator's isolated loopback to that collector
with adb reverse — the reverse-direction twin of the resident UI Automator server's forward_cmd
(BE-0245). The scenario schema, assertion pipeline, and capability model are all unchanged: network
stays a deliberately ungated construct the app-side collector satisfies without a backend advertising it,
exactly the accommodation idb already relies on.

What changed

• BajutsuAndroid/src/main/java/dev/bajutsu/android/BajutsuNet.kt (new) — the OkHttp interceptor: a
  no-op unless BAJUTSU_COLLECTOR is configured, POSTs each completed exchange as JSON matching
  NetworkExchange field-for-field over a separate reporting client. A report failure (a body-peek or
  JSON-build throw) is caught and logged rather than allowed to fail the app's real request.
• bajutsu/adb.py — reverse_cmd / reverse_remove_cmd, siblings of forward_cmd /
  forward_remove_cmd, tunneling the same device and host port so the injected collector URL resolves
  on-device unchanged.
• bajutsu/platform_lifecycle/protocols.py + environments/{android,ios,web}.py — a new
  RunEnvironment.bridge_collector(port) seam (a third "no-op implementation" decline idiom, documented
  alongside the existing two): AndroidEnvironment runs adb reverse and returns a real teardown thunk;
  iOS/web return a no-op, since the Simulator shares the host loopback and web observes via the driver.
• bajutsu/runner/pool.py — wires bridge_collector into the lease right before launch (alongside
  the existing BAJUTSU_COLLECTOR env injection) and tears it down on both release() and a launch
  failure, so one failed lease never leaks a tunnel.
• bajutsu/drivers/adb.py — a comment recording why AdbDriver.CAPABILITIES must not advertise
  network (that token means native driver observation), locked by a preflight test.
• Showcase reconciliation — compose/views Net.kt migrate from HttpURLConnection to OkHttp
  with BajutsuNet.interceptor() installed (debug-gated configure() call in MainActivity, matching
  the clipboard receiver's precedent); a new network_android.yaml scenario and a network (adb) job in
  android-e2e.yml prove a request assertion observes real emulator traffic end to end, via a
  throwaway host stub (net_stub.py) the emulator reaches at its 10.0.2.2 host-loopback alias.
• Docs — docs/network.md (+ ja), demos/showcase/SPEC.md (+ ja), and docs/architecture.md
  (+ ja) updated to describe Android's network capture; BajutsuAndroid/README.md (+ ja) gains a
  network-capture section.

Scope

Capture is bounded to OkHttp-originated HTTP(S) traffic — the same bounded shape as iOS's
URLSession-only scope. HttpURLConnection, other HTTP clients, and WebView traffic are out of scope
(a WebView needs its own follow-up, as it did on iOS, BE-0037). Deterministic mocks (stub
responses) are deliberately deferred to a follow-up item — this PR ships observation only, matching the
proposal's "Alternatives considered."

Prime-directive compliance

• AI never judges — no LLM anywhere on this path; the interceptor, collector, and request
  assertion are all deterministic.
• Determinism first — no fixed sleep; the on-device scenario waits on the mirrored *.status
  value and the collector via condition waits.
• App-agnostic — the interceptor is a library every app embeds identically; no per-app config.

Verification

make check
  ruff format/lint: All checks passed!
  mypy: Success: no issues found in 310 source files
  4369 passed, 3 skipped, coverage 92.39% (floor 89%)

New fast-gate coverage: reverse_cmd/reverse_remove_cmd command shapes, AndroidEnvironment.bridge_collector
establishing/tearing down the tunnel (including a teardown-failure-is-swallowed-and-logged case), the
pool wiring the bridge before launch and releasing it on both success and launch failure, the interceptor's
JSON wire-shape parsing into NetworkExchange, and that AdbDriver.CAPABILITIES excludes network while
a request scenario still preflight-cleans.

Not covered by make check (no JVM in the fast gate): the real Kotlin interceptor. It's proven by
the new network (adb) job in android-e2e.yml, an on-device CI run against a real emulator — not yet
executed in this session since it requires a booted AVD.

Reviewed with the built-in simplify/code-review discipline plus the pr-review-toolkit
silent-failure-hunter, type-design-analyzer, and pr-test-analyzer agents; their findings (an
unguarded report() call that could fail the app's real request, a swallowed-without-logging adb
failure, the Protocol docstring's declining-idiom list missing the new method, and an ungated configure()
call in the showcase) are all fixed in this diff.

🤖 Generated with Claude Code