ci: macOS runs serial + unready fast-fail + record/isolated docs (#190 follow-ups)

[borisbat]

Summary

Follow-ups from the #190 investigation. Three things, all stemming from how the headless suite behaves on the weak GitHub macOS runner.

1. macOS runs the suite serially (no --isolated-mode)

The headless suite was occasionally erroring with "daslang-live not ready after 30s" (first seen as a test_plot_getters_render flake). Root cause: the GitHub macOS runner is only ~3 vCPU, but the suite runs --isolated-mode-threads 4 — 4 parallel daslang-live hosts starve each other so hard that a host's /status health check takes ~10s to answer, racing the client's 10s per-request timeout. wait_until_ready never registers a success, burns its 30s budget, and the test errors before its body runs. It's not test-specific and not a logic bug — just CPU starvation.

Fix: macOS opts out of isolated mode (serial, one host at a time → no contention). ubuntu keeps isolated-4. The macOS sweep timeout is bumped to 1800s since serial is slower. Verified locally: full suite serial headless = 169/169 (198s).

2. Unready path fails fast (no 120s dead-wait)

The failure above was reported at 120s, not 30s: with_imgui_app's unready branch read the child to EOF but never told it to quit, so it rode to the 120s popen watchdog before surfacing the (already-decided-at-30s) "not ready" panic. Now it posts /shutdown first, so the host exits once it drains and the failure surfaces promptly. The 120s cap stays as the outer safety net (not lowered, so genuinely-slow test bodies aren't killed).

3. Docs: recordings can't run in macOS CI (no GL)

From the same investigation (the original #190 framing): the GitHub macOS runner has no usable OpenGL surface — a windowed recording host fails glfwCreateWindow with Glfw Error 65545: NSGL: Failed to find a suitable pixel format, renders zero frames, and that zero-frame output (not a render throttle) is what made content-time gating spin. Documented in tests.yml, alongside why isolated mode is correctness-safe (every [test] is with_imgui_app — headless, no GL window) and the rule that any future recording-bearing [test] must run in a separate, non-isolated step on a GL-capable runner.

Related

• daScript#3030 (merged): App Nap disable + recording-scoped glfwSwapInterval(0) + GL-backend log. Useful on real Macs; doesn't affect the CI case (no window).
• macOS CI: recording frame capture stalls during input-less holds (content-time gating degraded to wall-clock) #190 itself is closed (the recording-on-macOS-CI limitation is documented; not fixable without headless GL).

Rebased on master so #191 stays intact.

Refs #190

🤖 Generated with Claude Code