Decide whether the 8 macOS-skipped daemon-mode tests should be re-enabled (or the guard made explicit)

[StefanSteiner]

Decision needed

Eight daemon-mode integration tests in hyperdb-mcp/tests/daemon_tests.rs are skipped on macOS via:

#[cfg_attr(target_os = "macos", ignore = "flaky on macOS CI — daemon startup exceeds 150s timeout")]

They run on Linux and Windows CI — macOS is the only platform where they don't. This issue is to decide whether that guard is still warranted or should be lifted, rather than leaving it as a silent cfg_attr.

The eight:

• daemon_mode_engine_connects_to_shared_hyperd
• daemon_mode_two_engines_share_same_hyperd
• daemon_mode_persistent_database_file_survives_engine_drop
• daemon_mode_persistent_engine_data_is_queryable
• hyperd_monitor_detects_killed_hyperd_and_restarts
• client_report_triggers_restart_after_kill
• engine_recovers_after_hyperd_killed
• daemon_mode_ephemeral_database_cleaned_up_on_drop

Why it matters

These cover the daemon's core value: shared-hyperd reuse across engines, persistence across engine drops, crash detection and restart, and ephemeral cleanup. Until recently all eight were unconditionally #[ignore]d, so the daemon shipped resident-by-default for months with no automated crash-recovery coverage. Re-enabling them on Linux/Windows immediately turned CI red and exposed a real endpoint-publish-ordering bug (fixed in #286). So the coverage is load-bearing — and macOS is currently the one tier not getting it.

The evidence on both sides

For lifting the guard: on an Apple M3 Max these run in ~15–22s each, nowhere near the 150s budget. The reason string blames "daemon startup exceeds 150s timeout," which does not reproduce on a normal local macOS run.

For keeping it: the flakiness was observed specifically on shared/loaded CI runners. When macOS-ignored tests were forced to run under load (load average ~4.8, another build in progress), daemon_mode_engine_connects_to_shared_hyperd failed with "TestDaemon did not start within 150s". macOS GitHub-hosted runners are slower and more contended than the Linux ones, so a budget that is generous locally can still be tight there.

Options

1. Lift the guard, raise the budget. Re-enable on macOS with a longer or adaptive startup timeout (scale the 150s wait, or key it off observed CI slowness) so all three platforms get crash-recovery coverage. Risk: reintroduces macOS CI flakiness if the real problem is variance rather than the absolute budget.
2. Investigate the macOS startup cost first. Determine why daemon+hyperd startup can exceed 150s on macOS runners (cold binary, code-signing/quarantine checks on first exec, runner I/O) before changing the guard. Most likely to produce a durable fix.
3. Keep the guard as a deliberate tradeoff, but say so explicitly in the reason string — "skipped on macOS CI runners due to startup variance; covered on Linux and Windows" — so it reads as a decision, not a suspected-stale workaround.

Recommendation: option 2 then 1 — measure the macOS startup path before widening the budget, since a 150s timeout that is already 7× the local runtime suggests the failure is a stall, not a slow-but-completing start.

Related recent daemon-test reliability items

• fix(mcp): publish a restarted hyperd endpoint to daemon.json before STATUS #286 — the endpoint-publish-ordering bug these tests caught once re-enabled on Linux.
• Flaky: scan_all_refused_returns_freeport_base sees another test's fixture discovery record #300 — a separate shared-state flake in daemon_tests.rs (scan_all_refused_returns_freeport_base).
• Windows Named Pipe: verify DACL denies other users, and measure read-path perf for MCP workloads #302 — Windows Named Pipe verification/perf, once the daemon moves to IPC.

Provenance

Raised while auditing the ignored-test surface. Verified against main at 8760df2: the eight cfg_attr(target_os = "macos", ignore) guards are at daemon_tests.rs:1469–1718.