fix: NFR review findings — error-path hardening + protocol/contract test coverage

[tiantt]

Summary

Three rounds of fixes driven by a systematic NFR (non-functional requirements) review across security / observability / reliability / performance / test-quality dimensions.

Round 1 (27169e1)

• Remove dead duplicate POST /run_sse registration (debug leftover with print; the telemetry-instrumented handler registered first and moved to routes[0] is the live one — runtime behavior unchanged)
• Redact credential headers (authorization/token) before recording gen_ai.request.headers on spans in simple_app (they bypassed the logging RedactionFilter; mirrors agent_server middleware)
• ve_sign: thread signing scope through as parameters instead of mutating module globals (concurrent calls for different services/regions could sign with each other's scope); reuse a shared requests.Session for connection pooling; add golden-vector tests pinning the signing algorithm
• utils.misc.retry: exponential backoff with cap, explicit retry_on, documented idempotency contract; runtime polling gains mild backoff (3s → 10s cap)
• Stream pipeline step-log / runtime failure-log downloads to disk chunk-wise (flat peak memory)
• clean_env fixture snapshots/restores the full environment — raw os.environ writes can no longer leak across tests

Round 2 (a50edd6)

• Fix 3 order-dependent tests that only "passed" via leaked CLOUD_PROVIDER env (they never validated the global-config path they claimed to; also patch global_config_exists())
• /run_sse SSE error events no longer echo internal exception details to clients
• Cap server-provided Retry-After at 30s (+ tests)
• a2a executor wrapper: initialize result before try — executor failures previously got masked by UnboundLocalError in the finally-block telemetry call
• simple_app: run sync entrypoints via asyncio.to_thread (blocking user code no longer stalls the event loop)
• code_pipeline log export: bounded-pool concurrent downloads stitched in original order
• +49 tests: real-FastAPI route-table assembly (regression guard for the duplicate-route class), agentkit.apps public export surface, A2A JSON-RPC round-trip validated with a2a.types, MCP round-trip via fastmcp in-memory client, offline unit tests for the 5 previously untested SDK service clients

Round 3 (25d737c) — shadow-review blockers + cross-cutting majors

• cr: bound the instance-creation status poll with a 30-min deadline (was an unbounded while True — an instance stuck in Creating hung the CLI forever)
• sandbox exec: terminal websocket gains a 30s connect timeout + protocol-level ping keepalive (30s interval / 10s timeout) — a silently dead server is now detected instead of hanging the session until Ctrl-C (the app-level ping is server-initiated, so the client needed its own probe)
• apps: sensitive-header exclusion centralized in agentkit.apps.utils.SENSITIVE_HEADERS and extended with x-security-token / cookie / x-api-key / proxy-authorization / api-key / set-cookie — was two hand-synced two-entry sets plus one inline copy; X-Security-Token (STS credentials, see agentkit.auth._sigv4) previously leaked onto spans
• a2a telemetry: inline the user-id fallback instead of importing google-adk's private _get_user_id (closes the undeclared-dependency ImportError flagged in earlier rounds), wrap trace_a2a_agent with dont_throw so a telemetry failure in the finally block can no longer mask the business exception, and log execute failures with stack + context_id
• agent server /invoke: error frames no longer echo raw str(e) into the SSE payload (information leak + unescaped JSON); parity with /run_sse — generic message + error_type, json.dumps-built
• CI: add a test workflow (3.10/3.12 matrix, pytest + coverage report artifact — no --cov-fail-under yet: establish the baseline first, then ratchet); declare pytest / pytest-cov / google-adk in the dev dependency group; pytest + coverage config in pyproject

Test status

python3 -m pytest tests → 910 passed, 0 failed (branch started at 856 passed / 4 failed; holds after round 3).

Deliberately not addressed (separate decisions)

• CORS default allow_origins=["*"] + allow_credentials=True, and no default endpoint auth — behavior/product decisions, breaking for existing deployments
• Dependency CVEs reported by GitHub (4 critical / 21 high) — separate upgrade branch recommended
• base_service_client response validation can leak raw pydantic.ValidationError outside the ApiError hierarchy
• requirements.txt ↔ pyproject.toml dependency drift (each lists packages the other lacks; several core deps unpinned) — needs a single-source-of-truth decision
• Logging RedactionFilter's opaque-token regex over-redacts (request_id/trace_id/paths become ***), while the auth credential chain has no logging at all — both need a scoped design pass

Note on base branch

Stacked on fix/agentkit-latent-bugs (this work builds on the _signed_request/http_defaults engineering-standards changes there). When that branch merges, this PR retargets automatically.