fix(a2a): do not fail tasks on observation timeout

[RuniThomsen]

Bug Description

Inbound A2A message/send waits for the live Hermes session, then _await_reply writes TASK_STATE_FAILED with [agent did not reply in time] when the caller window (default 300s) elapses. The task is still running. Abel's board then archives the row. Timeout is not a terminal state (A2A §3.1.3).

Same-family as Seed #1412 (session-inbound never-ack → honest terminal). This cut is the caller observation seam on native a2a-platform, not Josh/Ib pickup-ack.

Root Cause

_await_reply used fut.result(timeout=max(0, remaining)). Past the deadline that is a 0s wait, then STATE_FAILED. The orphan watchdog could also fail a working task that still had a live pending waiter.

Fix

• After the observation window, keep waiting on the same future. Late complete() still lands.
• fail_orphans(..., skip_ids=live pending) so a live waiter is not "orphaned".
• Profile-CLI subprocess timeout is unchanged (that process was killed).

How to Verify

python -m pytest tests/plugins/test_a2a_plugin.py::TestObservationTimeoutNotFailed tests/plugins/test_a2a_phase23.py::TestTaskStore::test_fail_orphans -v

Sabotage: restore timeout=max(0.0, remaining) plus the STATE_FAILED / [agent did not reply in time] return → the late-complete test is RED; restore GREEN.

Test Plan

• Late reply after deadline → TASK_STATE_COMPLETED, not failed
• Watchdog skip live pending ids
• Existing fail_orphans still fails true orphans
• test_a2a_plugin.py + test_a2a_phase23.py 153 passed

Risk Assessment

Low–medium. Holding message/send threads now wait until the agent actually replies (or a real failure / client disconnect). Callers that already treat HTTP timeout as local observation keep GetTask as working and can receive a late push. PR ≠ merge ≠ live gateway (Wolf gateway not restarted).

Exclusions

• No Seed bump (that is PR 800 / 7.6.20 pickup-ack).
• No Kenneth-box / :3979 bounce, no board file write.

[Enough1122]

AI code review — automated review for reference; please use your judgment.

fix(a2a): do not fail tasks on observation timeout (A2A 3.1.3) — correctly separates caller observation window from task terminal state.

• plugins/platforms/a2a/adapter.py:38 _await_reply — replaces immediate STATE_FAILED/[agent did not reply in time] on deadline with one-time logger.info observation window elapsed then loops on fut.result(timeout=_SSE_KEEPALIVE) (or remaining window). Only keepalive() raise or exception maps to STATE_FAILED; late answer still returns STATE_COMPLETED. Preserves keepalive semantics for SSE.
• plugins/platforms/a2a/adapter.py:25 watchdog path — snapshot _pending under lock then fail_orphans(skip_ids=live); protocol.py:100 adds skip_ids filter so live waiter ids never fail even if created_at is stale. Orphan watchdog still fails detached tasks after 5 min.
• plugins/platforms/a2a/DESIGN.md:12 documents non-terminal observation vs orphan distinction.
• Tests test_await_reply_keeps_waiting_after_deadline_then_completes and test_watchdog_skips_tasks_with_live_waiters cover late answer and live-skip.

Approve.