[Zhoujie628]

Closes #12

Found while auditing this engine against the Java engine (workflow-engine-sdk-java) and against that engine's commit history.

Scheduling correctness

A merge step could start before its own predecessor had run. Step readiness was computed from direct predecessors and then filtered to those already activated, so a graph whose branches activate at different times let the merge consume a partial input and still report success. Readiness now considers every active ancestor, and the deadlock path reports the missing steps instead of returning a silent result.

Protocol and transport, aligned with the Java engine

• Recover an interrupted SSE stream by polling the task, bounded by the new task_poll_interval_seconds (default 20s, floor 0.1s). The poll runs only after an interruption, never while the stream is alive.
• Open a negotiation for a bare Propose message, and fail loudly when a negotiation round yields no task instead of sleeping in a spin.
• Bound the notification acknowledgement with the new notification_ack_timeout_seconds (default 300s), so a remote that accepts the stream but never acknowledges cannot leave the caller waiting forever.
• Observe transport activity per body chunk, including an SSE comment heartbeat that never becomes a decoded event, and judge notification liveness from it. A subscription receiving only heartbeats was previously reported as dead, and a caller acting on that tore down a live stream; a genuinely dead stream leaked its connection and task forever. A subscription is now released after send_timeout_seconds without any transport activity.
• Report the last business event separately through last_business_event_at, keeping last_event_at as transport activity.

Security

• Match sensitive header names after stripping - and _, so api_key, x_api_key, access_session, pwd and passwd are redacted like their other spellings. A header registered through register_sensitive_header now matches in any spelling.
• Fail closed on TLS material for the registry path, and reject .env keys that are not valid environment-variable names.

Correctness and cleanup

Fix the repeated-argument parse patch, cache one A2A client per agent and invalidate the cache when agent cards change, forward task_id in the intermediate status event, bound stream_message by timeout_seconds, accept a custom CA, mTLS identity and CRL for the registry helpers, and make workflow_engine.__version__ the single source of the distribution version.

Behavior changes to review

Several changes are observable to existing callers:

• NotificationHeartbeat.last_event_at now reports transport activity rather than the last business event, and is_healthy() judges liveness from transport activity.
• A notification stream is released after send_timeout_seconds without transport activity. Hosts that relied on a subscription staying open forever must be updated.
• stream_message is bounded by timeout_seconds, which defaults to the interaction deadline. Long-lived streams need an explicit larger value.
• Merge steps now wait for every active ancestor, so such a step may start one round later than before.

Verification

120 tests pass, 72 of them new across test_review_fixes.py, test_p0_alignment.py and test_java_history_parity.py. verify_sdk.py reports VERIFICATION: PASS.