test: fill A2A adapter unit test gaps

[AlexanderZ-Band]

Summary

Part 1 — unit test gaps:

• Adds coverage for 10 undertested paths across the outbound A2AAdapter and A2AGatewayAdapter: peer-not-found resolution, peer-list pagination, artifact-overwrite reduction, TASK_STATE_AUTH_REQUIRED handling, and several gateway JSON-RPC route edge cases (malformed body, non-scalar request id, SendStreamingMessage, missing-id task methods, partial room-creation state).
• Two of these gaps turned out to be real bugs, surfaced while writing their regression tests, and fixed alongside the new coverage:
  • A2AGatewayAdapter._execute_a2a: a REST failure while posting to Band left the remote A2A caller stuck on a WORKING task forever, since no terminal FAILED event was ever published on that path. Now calls pending.fail(...) before re-raising.
  • A2AAdapter.cleanup_all: an exception from the A2A client's close() skipped closing the owned httpx transport, leaking it on shutdown. Now closes the transport in a finally.

Part 2 — live E2E smokes:

• Three bespoke baseline smokes under tests/e2e/baseline/smoke/adapters/ (A2A is a protocol bridge, not a matrix adapter, so these follow the test_parlant.py pattern):
  • test_a2a.py — a live A2AAdapter against a new scripted, non-Band A2A counterparty fixture (a2aServer.py, built on a2a-sdk's own server primitives — no LLM key needed). Covers both a canned-reply happy path and a scripted remote task failure.
  • test_a2a_gateway.py — the official a2a-sdk reference client (a2a.client.Client) driving a live A2AGatewayAdapter that exposes a real Anthropic-backed Band peer — validates the gateway's JSON-RPC server independent of our own A2AAdapter.
  • test_a2a_roundtrip.py — the full Band-to-Band shape: Band Agent A (A2AAdapter) → live gateway → Band Agent B (Anthropic) → reply relayed all the way back.
• Also fixes a race surfaced while building these: GatewayServer.start() returned as soon as serve() was scheduled, before uvicorn was actually listening — a caller dialing in immediately after (a real A2A client, or these new smokes) could hit connection-refused. start() now waits for uvicorn's ready signal, bounded by a timeout, with a small regression test.
• All three smokes were subsequently run live against a real platform + ANTHROPIC_API_KEY (see Part 3 — two of the three initially failed live and led to two more real bug fixes); all four tests (test_a2a.py has two) now pass live with zero reruns.

Part 3 — two more real bugs found running the smokes live:

• A2AAdapter's owned httpx client had no read-timeout override, so httpx's default 5s read timeout fired on the ordinary multi-second gap between SSE events during a real remote turn (a live LLM call), not just a genuine hang — surfaced as a spurious A2AClientTimeoutError on any real turn slower than 5s. Fixed by leaving read unbounded while keeping connect/write/pool bounded, so a genuinely dead peer still fails promptly. Same fix applied to the raw a2a-sdk client built in test_a2a_gateway.py.
• The A2A gateway leaked sse_starlette's process-global shutdown flag across instances. sse_starlette's background shutdown watcher polls whichever uvicorn.Server owns the process's SIGTERM slot and promotes its should_exit to the process-global AppStatus.should_exit. GatewayServer.stop() sets should_exit directly (never via a real signal) — so once any gateway in a process stops, every later GatewayServer's SSE streams in that same process get cancelled immediately ("ASGI callable returned without completing response.", an incomplete chunked read on the client side). This is exactly why test_a2a_roundtrip.py failed right after test_a2a_gateway.py passed in the same pytest session — each test builds its own GatewayServer. band.integrations.mcp.local_server had already hit and fixed the identical bug for itself (a real Windows CI hang); the A2A gateway now disables the same process-global switch directly, rather than depending on that other module happening to be imported first. Added a regression test with two real GatewayServer instances in one process (confirmed it reproduces the exact same failure signature with the fix reverted, and passes with it restored).

Closes INT-1357.

Test plan

• uv run pytest tests/integrations/a2a/ tests/integrations/mcp/ -v --no-cov (89 passed)
• uv run pytest tests/ --ignore=tests/integration/ --ignore=tests/e2e/ -q (5288 passed, 123 skipped, 0 failed)
• uv run ruff check . / uv run ruff format .
• uv run pyrefly check (0 errors)
• E2E_TESTS_ENABLED=true uv run pytest tests/e2e/baseline/smoke/adapters/test_a2a.py tests/e2e/baseline/smoke/adapters/test_a2a_gateway.py tests/e2e/baseline/smoke/adapters/test_a2a_roundtrip.py -v -s --no-cov against a live platform + ANTHROPIC_API_KEY — all 4 tests pass, zero reruns (117s total)
• Verified the gateway shutdown-leak regression test actually catches the bug: reverted the fix, reran, got the identical live-failure signature (RemoteProtocolError: incomplete chunked read / ASGI callable returned without completing response.); restored the fix and confirmed it passes again

🤖 Generated with Claude Code

https://claude.ai/code/session_017iDgB8tuW6tPom4PFomMzr

[linear-code]

INT-1357

[amit-gazal-band]

Flagging three issues from review, focused on the SSE-timeout fix and the new gateway startup-wait logic.