feat: isolate coding agent processes by room

[AlexanderZ-Band]

Summary

• assign each room an absolute workspace via workspace_for_room(room_id), defaulting to an isolated ./.band-workspaces/<room-id> directory when unset
• lazily create and own one local stdio coding-agent process and protocol state per room
• pass the resolved room workspace to Codex thread/start, not a shared cwd
• reject remote, shared, and injected transports that cannot preserve room isolation
• enforce one live room per resolved workspace, including symlink-equivalent paths
• make workspace/runtime creation atomic; clean up failed Codex client startup and evict the failed client
• remove Codex's shared fallback state and update ACP/Codex documentation and examples

Validation

• uv run pytest tests/adapters/test_room_workspace_isolation.py -q — 4 passed
• uv run ruff check src/band/adapters/codex.py src/band/integrations/acp/client_adapter.py src/band/workspaces.py tests/adapters/test_room_workspace_isolation.py examples/codex/01_basic_agent.py examples/codex/02_tom_agent.py examples/codex/03_jerry_agent.py examples/run_agent.py examples/docker_demo/agents/dev/main.py
• uv run pyrefly check src/band/adapters/codex.py src/band/integrations/acp/client_adapter.py src/band/workspaces.py tests/adapters/test_room_workspace_isolation.py
• python -m compileall for changed examples
• git diff --check

Draft gate

Legacy ACP/Codex fixture suites (tests/adapters/test_codex_adapter.py, tests/integrations/acp/test_client_adapter.py, tests/integrations/acp/test_client_adapter_behavior.py) construct these adapters against the removed shared-process/injected-client APIs (cwd=, client_factory=, spawn_process=) and fail without a workspace-lifecycle migration: 143 failed / 58 passed in the Codex suite, 207 failed / 88 passed / 14 errors in the ACP suite. This PR stays draft until those fixtures move to room-owned stdio lifecycle equivalents.

[linear-code]

INT-1428

[AlexanderZ-Band]

Review pass: fixed room-workspace ownership, removal of shared Codex fallback state, failed Codex startup cleanup, stale transport documentation, and focused regression coverage. The legacy ACP/Codex fixture suite still targets the intentionally removed shared-process and injected-client APIs; it remains the reason this PR stays draft.