test(cli): restore the run-command process-boundary contracts

[UncertaintyDeterminesYou4ndMe]

Summary

Restores the six process-boundary contracts that were retained on a real subprocess in #2476 and lost when 802855c removed the process fixture — as invited in the closing review of #2476 (Refs #2476, refs #2387).

The injected MakaRunDeps seam covers the semantics of stdinIsTTY/readStdin, but not the boundary itself: a real pipe feeding the prompt, a SIGINT delivered by the operating system observed as exit 130 with nothing on stdout, and the fail-closed sandbox path traversing a real child process. Current main asserts none of these (stdinIsTTY/readStdin are exercised in process only; no test anywhere asserts SIGINT → 130; the remaining spawn in cli.test.ts covers profile/identity files, not the run path).

• run-command-fixture.ts (new) — a subprocess entry over today's runMakaTextCliCore with default environment deps (real stdin, real SIGINT handler, real exit code) and a minimal scripted adapter: four scenarios (echo, sandbox-boundary, slow, graph-wait), no env-var assertion matrix, none of the in-process duplication that 802855c removed. fixture-ready is written only from inside the run (after the core has installed its SIGINT handler), so the test's signal cannot race the default disposition.

• run-command.test.ts — one describe('maka run process contract') block with the six tests, assertions unchanged from the pre-deletion suite:

  Contract	Tests
  Piped non-TTY stdin (run -, implicit stdin prompt, positional + stdin context)	3
  SIGINT → exit 130, empty stdout through a real pipe (plain run; during Graph completion wait)	2
  Fail-closed sandbox boundary reaching a non-interactive run	1

  Both signal tests carry bounded SIGKILL guards (never-ready, never-exiting) so a regression fails the suite instead of hanging it.

No production changes.

Verification

• packages/cli: node --test dist/__tests__/run-command.test.js — 11/11, three consecutive rounds (the six subprocess tests add ~0.8s total); full CLI suite 442/442.
• Fault injection (in dist/, reverted): disabling if (interrupted) return 130 fails exactly the two SIGINT tests; disabling the implicit stdin read fails exactly the two non-run - stdin tests.
• Biome clean on both files; ASF header audit clean; tsc clean.
• Not run: workspaces this change does not touch.

AI use

Select exactly one:

• No generative tool made a substantive contribution
• Generative tooling made a substantive contribution

Tool(s) and scope: Claude Code authored the fixture and tests (ported from the pre-deletion suite) and ran the verification; commits carry Generated-by: Claude Code trailers. Reviewed and submitted by a human.

Checklist

• Tests cover the change and fail without it
• Lint, format, typecheck and the affected suites pass locally

Does this PR entail a change in behavior?

• Yes — described under Summary above
• No

[UncertaintyDeterminesYou4ndMe]

The failing check is Check ASF source headers, and the five files it names are all from d807c0078 (#3250) on current main — none are touched by this PR (both files here carry the header; the audit passes locally against this branch's own changes). The in-flight fix is #3708; once it lands I will rebase to re-run CI on a base where the audit is green.

[Astro-Han]

Adds a 157-line subprocess entry and four scripted scenarios covering three maka run contracts that only hold across a real process boundary. SIGINT currently appears nowhere in the repo's tests.

No P-graded issues. I checked why these were removed originally: duplication and fake-only coverage, not flakiness. Fault injection failed in exactly the right cases both times.

Two ungraded notes. sandbox-boundary is entirely within the seam and doesn't need a subprocess, and the three stdin scenarios compress to one — the two SIGINT cases carry nearly all the value, and ~160 lines would do. More substantively, this exercises a synthetic process boundary: it bypasses production's resolveMakaCliExitCode and the exit grace timer, so it can't catch a regression in either.

---

AI-assisted review: a Claude Code subagent analysed the change; I verified the deletion rationale from history and the fault-injection results myself. No tests run. AI review is not independent human review.