fix(omo-senpi): prevent orphaned memory test fixtures

[code-yeongyu]

Root cause

Memory lock fixtures and the model-preflight pipe-holder regression harness could outlive an abruptly terminated test parent. The lock fixture's lifetime was not fully event-driven, and the preflight wrapper/grandchild used forever-running intervals. Facts-test teardown also removed temporary roots without awaiting tracked child termination. Numeric PID cleanup could signal a reused, unrelated process.

Changes

• Bind hold-lock.ts to the parent-owned stdin channel (end/close/error) with no polling interval.
• Replace model-preflight wrapper/grandchild intervals with a parent-owned loopback control socket; closing it tears down both helpers.
• Add bounded liveness helpers with zombie-aware checks and command identity snapshots; cleanup signals only a still-matching PID.
• Make facts fixture teardown ordered and bounded: SIGTERM, await exit, SIGKILL escalation, await again, then remove roots.
• Add real-process lifecycle regressions while preserving the full pipe-holder degradation sensitivity.

Test evidence

• Focused lifecycle + model-preflight: 12 pass, 0 fail.
• Facts pruning + failure streaks: 25 pass, 0 fail.
• bunx tsgo --noEmit -p packages/omo-senpi/tsconfig.json: exit 0.
• git diff --check: pass.
• Full bun run test:senpi was started and produced extensive passing output but exceeded the 300s execution cap before completion; not claimed as passing.
• Baseline replay against origin/dev confirmed the old model-preflight interval harness remained present; the borrowed stdin fixture regression is not a valid RED discriminator under Bun/macOS because Bun closes inherited stdin when the wrapper exits.

Supersedes #7342.

---

Summary by cubic

Prevents omo-senpi memory lock fixtures and preflight pipe-holder helpers from outliving an abruptly terminated test parent, which previously stranded orphaned processes holding lock markers near 100% CPU.

• hold-lock.ts exits when its parent-owned stdin channel reports end, close, or error, replacing the PPID polling watchdog, and destroys stdin on every ownership-loss path so the process can terminate.
• The model-preflight wrapper and grandchild park on a parent-owned loopback control socket; closing it tears down both helpers.
• Facts fixture teardown is ordered and bounded: SIGTERM, await exit, SIGKILL escalation, await again, then remove temp roots.
• Lifecycle regression tests observe the original helpers through pid files and exit markers, asserting the wrapper dies on probe timeout, the grandchild survives holding the pipes, and neither outlives control-channel teardown.
• Liveness helpers are zombie-aware, verify PID command identity before signalling, and detect process state portably across Linux and macOS; pidTerminalWithin rides the same bounded probe on every platform because a zombie keeps its /proc entry and can sleep through a watcher.
• Foreign-process and file waits use bounded condition probes (probeUntil) because fs.watch file-creation events are unreliable on some platforms and caused CI timeouts.
• The model-preflight grandchild test awaits the grandchild's connected marker before closing the control channel and uses a 30s budget so Bun's 5s default can't cut off wrapper plus grandchild spawn on slow runners.

^{Written for commit a9f210d. Summary will update on new commits.}