[elliotBraem]

What

Two things from the process-death regression investigation:

1. Boot-probe wedge fix (probeHttpOk)

The dev session's readiness probe awaited fetchRaw — a promise bridge around a nested Effect.runPromise — from inside the probe fiber. The nested runtime intermittently wedged (repro'd 5/5 clean, 4/4 pass with a single stderr write shifting timing), stalling boots until the ~90s probe deadline self-healed. This is a plausible contributor to slow CI boots in the framework regression stats.

• probeHttpOk is now Effect-native (fetchWithRetryEff + orElseSucceed(false)) — same deadline, same all-failures→false semantics, no nested runtime.
• The host-manifest fetch (same nested pattern, remote-host boot path) got the same treatment with fail-loud status/validation preserved.
• The graceful-overrun force backstop (armForce/disarmForce) is now a forked Effect.sleep fiber via Effect.forkDetach (a scoped fork inside a finalizer attaches to the closing scope and dies with it) instead of raw setTimeout — the last non-Effect resource management on the teardown path.

2. Unit process-death harness

tests/unit/dev-session-processes.test.ts drives the real runDevSession against real detached group-leader children and watcher grandchildren (30 MB fixture node servers instead of the 5–6 GB full stack):

• graceful shutdown kills children + grandchildren, frees ports, cleans the registry
• a child ignoring SIGTERM is escalated to SIGKILL by teardown
• double-graceful escalates straight to force
• orphan watchdog force-exits on reparenting (the deps.ppid seam)
• startup adoption reaps a dead session's live orphans and prunes the registry

This complements the existing mock-based dev-session.test.ts (session semantics) and process-kill.test.ts (kill primitives) with the real-topology layer between them and the slow Go suite.

Verification

• New suite + all affected suites: 30/30 green locally (~35s)
• pnpm run lint and pnpm run typecheck pass
• CI runs the new file automatically via framework-tests (pnpm --filter everything-dev test, gated on packages/everything-dev/ changes)

Follow-ups (next PRs)

• Confined stacks in the Go framework suite: most tests boot a fixture bos.app.ts with the fake services instead of all 9 real workspaces; keep one full-stack boot (real rspack watchers) as the safety net. Then CI slots can go from 1 back to 2–3 and the 35-min timeout drops (the RAM-saturation rationale is gone).

[elliotBraem]

Stacked Phase 2 (310e1e0): every framework test but one now boots the confined death-contract fixture (minimal authored config, five fake node services) instead of the full 9-child stack. Orphan test: 213s → 4.6s. One full-stack boot (sigterm) remains as the real-rspack-watcher safety net. CI: 2 stack slots instead of serial, timeout 35 → 12 min. Lint/typecheck green; Go vet clean.