fix(delegation): runtime handles 202+queued; canvas surfaces delegation rows

[HongmingWang-Rabbit]

What this fixes

Two bugs that compounded into the "Director does the work itself" UX you saw today (workspace `8b0bf0b3`):

1. Runtime treated 202+queued as silent failure

`workspace/builtin_tools/delegation.py:_execute_delegation` only handled HTTP 200. When the peer's a2a-proxy returned HTTP 202 + `{queued: true}` (single-SDK-session bottleneck on the peer), the loop fell through. Two iterations later `if "error" in result` tried to access an unbound `result`, the goroutine ended quietly, the delegation stayed at FAILED with `error="None"`. The LLM checking status saw "failed" + the platform's "Delegation queued — target at capacity" log line in chat context, concluded the peer was permanently unavailable, and bypassed delegation to do the work itself.

DB evidence from the live workspace (6 delegation rows, all `status='queued'`):

```
delegation queued → UX Researcher "Delegating to f6f3a023..."
delegation queued → UX Researcher "Delegation queued — target at capacity"
delegation queued → Visual Designer "Delegating to a2e92524..."
delegation queued → Visual Designer "Delegation queued — target at capacity"
delegation queued → React Engineer "Delegating to fd1a1f96..."
delegation queued → React Engineer "Delegation queued — target at capacity"
```

2. Canvas Agent Comms tab dropped every delegation row

`AgentCommsPanel.tsx` filtered to `a2a_send | a2a_receive` only. `activity_type='delegation'` rows never reached `toCommMessage`. User saw "No agent-to-agent communications yet" while the DB had 6+ delegations.

How it's fixed

File	Change
`workspace/builtin_tools/delegation.py`	Add `DelegationStatus.QUEUED`. Explicit `if a2a_resp.status_code == 202` branch in `_execute_delegation` — recognizes `{queued: true}`, marks local state QUEUED, mirrors to platform, returns cleanly without retrying. `check_delegation_status` docstring extended with explicit per-status guidance: "queued → wait, peer is on prior task, do NOT bypass."
`canvas/src/components/tabs/chat/AgentCommsPanel.tsx`	Include `delegation` in both the initial filter and the WS push filter. New `delegation` branch in `toCommMessage` — maps as outbound (always; platform proxies on our behalf), uses `summary` as primary text.

Tests

• 3 new Python tests (`workspace/tests/test_delegation.py` → `TestA2AQueued`):
  • 202+queued marks status as QUEUED, no error field
  • 202+queued does NOT retry the A2A POST (counted by URL match — mock is shared across all AsyncClient calls)
  • bare 202 without `{queued:true}` still falls through to existing retry-then-FAILED path
• 3 new TS tests (`AgentCommsPanel.test.ts`):
  • `delegate` row maps as outbound with summary text
  • `delegate_result` queued row preserves `status='queued'` (load-bearing for the LLM's wait-vs-bypass decision)
  • missing `target_id` returns `null` instead of rendering a ghost

All 20 `test_delegation.py` tests pass; all 8 `AgentCommsPanel.test.ts` tests pass.

What this PR does NOT solve

The underlying single-SDK-session bottleneck (peer can only handle one A2A request at a time). Tracked as task #102 — real architectural work. This PR makes the runtime handle the queueing correctly so the LLM doesn't bail, and makes delegations visible in Agent Comms so operators can see what's happening. Once #102 lands, the queueing will become rare anyway.

Test plan

• `pytest workspace/tests/test_delegation.py` green (20/20)
• `vitest run AgentCommsPanel.test.ts` green (8/8)
• Manual: provision a director + 3 sub-agents, send director a multi-step task, observe Agent Comms panel populates with delegation rows including queued ones; observe director NOT bypassing when peer is queued (status surfaced as "queued" in chat, director acknowledges and waits)

🤖 Generated with Claude Code

[HongmingWang-Rabbit]

Still skipping: failed E2E correlates with the active Cloudflare Minor Service Outage (status page indicator updated 22:08:40Z). CP #284's fast-fail surfaces CF API failures in ~16s instead of 900s — that's working as intended; the underlying CF outage is the blocker. Re-trigger when cloudflarestatus.com returns to operational.

[HongmingWang-Rabbit]

Code review (5-axis)

Correctness: ✓ Two-part fix lands cleanly:

1. Runtime side (delegation.py) — explicit 202+queued branch returns early after marking DelegationStatus.QUEUED, calling _notify_completion(..., 'queued'), and _update_delegation_on_platform(..., 'queued', ...). The malformed-JSON guard (try: a2a_resp.json() except: queued_body = {}) plus the strict is True check on queued correctly falls through to the existing failure path on 202-without-queued-flag — no silent state confusion.

2. Canvas side (AgentCommsPanel) — filter expanded + toCommMessage mapping for activity_type === 'delegation'. The !peerId guard returns null for malformed rows (matches the test contract).

Tests: ✓ Coverage is the right shape — TestA2AQueued.test_queued_marks_status_queued_not_failed AND test_queued_does_not_retry. The retry-count assertion is non-trivial (filtering A2A POSTs from platform-sync POSTs by URL match) but correct. Canvas tests cover delegate / delegate_result + the null-target_id defensive case.

Architecture: ✓ QUEUED enum addition is forward-compatible (existing clients see other values unchanged). Status-semantics docstring on check_delegation_status is the right place to teach the LLM what queued means — that's where the prompt-side guidance has to live.

Security: N/A — no auth/data-boundary changes; the queued-body is parsed but only .get('queued') is read, so a hostile peer can't smuggle anything via the rest of the JSON.

Performance: ✓ One additional branch + one body parse on 202 responses. Negligible.

FYI

• The queued_body malformed-JSON path silently sets {} and falls through. Worth a one-liner log here too (mirrors the env-var pattern in fix(a2a-proxy): close 60s context-canceled gap on long silent runs #2128) — would help diagnose a future "peer is returning weird 202s" incident. Not blocking.
• The LLM-prompt-side "don't bypass on queued" guidance lives in the docstring. Whether the runtime actually honors it depends on the agent's prompt construction — outside this PR's scope, but worth a follow-up to verify the LLM does the right thing in a real queued scenario before declaring incident-closed.

LGTM. Auto-merge already armed.