fix(agents): wire persistent teams and repair worker lifecycles

[ericleepi314]

Completed background workers previously could appear to resume without starting a model loop, and the team tools had no production path for persistent teammate execution. This change connects those lifecycles end to end: workers process follow-ups with their history, teammates remain available for assignments, and explicit messages reach the correct peer or leader.

The implementation also:

• Shares admission, progress, transcript persistence, session/parent notification routing, and bounded shutdown across local, team, and workflow workers.
• Implements leader/member identity, mailbox consumption, dependency-aware task pickup, completion-hook veto, matching plan/shutdown decisions, and interruption of pending permission requests through the WebSocket server.
• Fixes concurrent name claims, late-message completion races, workflow startup/stop races, budget accounting and replay checkpoints.
• Honors Git worktree isolation, fails when isolation cannot be created, preserves edits/commits, and retains a checkout still used by a background descendant.
• Repairs four existing Windows test-fixture assumptions (UTF-8, literal CRLF, unreadable-directory setup, and native home-directory environment) identified in the prior main-branch CI run. Assertions are preserved.

Validation and the comparison with the Claude Code reference are documented in docs/multi-agent-runtime-verification.md. New end-to-end tests execute real query loops, tools, files, registries and WebSocket connections with a scripted external provider. A separate live DeepSeek smoke run passed background Read, same-ID resumption, peer-to-leader delivery, and approved shutdown/deletion.

Validation:

• Final GitHub Actions run: all 8 checks passed, including Python on Linux (10,755 passed; 19 skipped) and Windows (10,725 passed; 48 skipped), desktop on both platforms, web and Harbor.
• Full Python suite: 10,758 passed, 16 skipped, 340 passing subtests (316.14 s, final revision).
• Portable Windows fixture group: 144 passed locally.
• Bridge/error-propagation regression: 51 passed. Heartbeat assertions await the persisted refresh; the mocked connection-error test retains retries without real network-backoff sleeps.
• Team/task regression: 163 passed, with 20 independent passes of the automatic-claim/durable-snapshot trace.
• Thread-aware admission and end-to-end regression: 54 passed. Admission tests use elapsed-time waits, thread-safe events, and teardown joins; the admission suite also passes with asyncio debug enabled.
• TUI typecheck and all 65 agent-tree/task-output tests pass.
• Black/isort applied; git diff --check passes. Four new runtime modules pass targeted mypy. Full mypy has 395 baseline diagnostics vs. 397 before, with zero added diagnostics after line-number normalization.
• The unchanged TUI's broader suite has 1,896 passed, 8 failed, 4 skipped. Seven failures reproduce on the untouched starting commit; the cursor-layout timeout passes on isolated rerun. These pre-existing display/configuration failures are documented, not reported as green.

Scope: persistent teams are in-process, and local worker continuation is session-owned. Remote/pane/UDS backends and automatic crash recovery are not included. Observed token budgets gate new work; already-running requests can exceed the threshold. These limits and stale-roster handling are documented.

[github-actions]

Test Results

     5 files   1 021 suites   20m 24s ⏱️
16 031 tests 16 009 ✅ 22 💤 0 ❌
32 032 runs  31 961 ✅ 71 💤 0 ❌

Results for commit 99d6c49.

♻️ This comment has been updated with latest results.