Let daemon sleep with idle MCP wrappers

[Marsu6996]

Summary

This PR fixes a daemon sleep regression where idle MCP wrappers/control-plane probes could keep the daemon awake indefinitely, and adds an on-demand wake path for launchd-managed daemons after sleep/hibernation.

Problem

Long-lived MCP sessions keep wrapper heartbeat files fresh, while watchdog/status probes arrive frequently. The daemon was treating those presence/control-plane signals as real activity, so the idle cycle could stay in WAKE and repeatedly do expensive wake-time work even when no memory operation was running.

There was also a recall hot path where the max-node community gate could scan the full assignment instead of only the current recall candidates.

Changes

• Count only real memory activity for daemon idle decisions:
  • active/pending drains
  • recent JSON-RPC tool traffic
  • not wrapper heartbeat freshness or watchdog/status probes
• Wake the managed daemon from the MCP wrapper on tools/call when using the default macOS/Windows endpoint:
  • write wake.signal before kickstart
  • wait briefly for socket readiness
  • skip custom sockets and IAI_MCP_DISABLE_DAEMON_AUTOSTART=1
• Allow true WAKE transitions to bypass S2 anti-oscillation.
• Keep RECALL_CONCURRENCY_DEFAULT=2 for multi-agent setups.
• Reduce recall CPU by scoring max-node community gates over current candidate embeddings only.
• Guard recall scoring/gating against NaN/Inf embeddings.
• Make daemon status prefer canonical lifecycle_state.json over stale legacy in-memory state, so daemon status does not report WAKE while the lifecycle FSM is DROWSY/SLEEP.

Validation

Passed:

• npm run typecheck && npm test in mcp-wrapper — 51 passed
• ruff check src/iai_mcp/concurrency.py src/iai_mcp/pipeline.py src/iai_mcp/socket_server.py tests/test_concurrency_session_open.py tests/test_mosaic_max_node_gate.py tests/test_recall_shared_cosine.py tests/test_socket_server_dispatch.py
• targeted pytest recall/sleep/socket/cascade/status suite — 51 passed in the final de-duplicated pass; earlier broader targeted pass was 130 passed, 8 skipped before removing the fix(daemon): snapshot state before threaded save to end tick crash #47 overlap from this branch

Live local smoke after installing the patch over the active daemon runtime:

• MCP wrapper initialized and listed 13 tools
• memory_recall returned non-error results
• canonical lifecycle state moved to DROWSY after the idle threshold
• daemon CPU dropped back to near-idle after the DROWSY edge work

Full local pytest was also run earlier. It reached 3640 passed, 122 skipped, 1 xfailed, with 2 local/environment failures unrelated to this diff:

• tests/test_mosaicsigma_native_import.py::test_both_submodules_share_one_so because the active site-packages directory contains several extension modules, not just iai_mcp_native.
• tests/test_watchdog_phys_footprint.py::test_phys_footprint_not_above_rss_on_clean_process under macOS memory-pressure/measurement conditions.

Review

Reviewed with an aggressive Claude pass. Initial blockers were addressed:

• wake after hibernation is now guaranteed from the MCP wrapper hot path, not only at session start;
• the silent recall pool cap was removed;
• recall concurrency default remains 2;
• the status finishing pass is covered by targeted tests.

Relationship with other open PRs

• fix(daemon): snapshot state before threaded save to end tick crash #47 is our separate focused PR for the threaded save_state snapshot race. This PR no longer includes that fix.
• Windows: fix daemon liveness, crypto key corruption, and auth-token resilience #44, Fix lifecycle-lock acquire() race that allowed duplicate daemons #46, and Windows: fix daemon wedge from blocking flock in hippo lock conversion #49 are separate Windows/lifecycle-lock/hippo-lock fixes and are not duplicates of this PR.

Operational notes

• The first tool call after hibernation may wait up to 7 seconds for the daemon socket to return before falling through to existing fallback/error paths.
• The wrapper wake path is intentionally limited to the default managed user endpoint; custom socket deployments own their own supervision.

Attribution

Designed by Marsu — Refined by Codex.

Co-Authored-By: Codex codex@openai.com

CI follow-up

• macOS CI failed once on tests/test_runtime_graph_centrality_child.py::test_parent_rss_lower_with_centrality_child with in_parent_delta=63.3MB and isolated_delta=65.0MB.
• Local reruns showed the intended signal clearly (168.5MB/49.4MB, 167.1MB/47.9MB), so the failure was a strict RSS allocator-noise assertion, not a runtime regression.
• The test now xfails only when RSS proof is inconclusive within a 32 MiB allocator-noise margin; it still fails if the child path is materially worse.