chore(createPR): remove subprocess + gadget timeouts so long pre-push hooks survive

[zbigniewsobiecki]

Summary

• Drop the three stacked timeouts on CreatePR so pre-push hooks running full test suites (5+ min) stop getting SIGTERM'd. Harness already budgets long tool calls — no value in a second shorter cap here.
• git commit + git push now call runCommand with wallTimeoutMs: 0, idleTimeoutMs: 0 (disabled, per runCommand's docstring convention).
• The CreatePR gadget's outer timeoutMs: 240000 is now 0, which llmist treats as "no outer timer" (if (timeoutMs && timeoutMs > 0) in its dispatch path).
• Heartbeat stays default (30 s stderr pulse) so operators still see [git-push] still running (Ns) ticks.

Why

Three wall-clocks used to stack on one code path, any of which could fire first:

Layer	Old cap	Result
git commit subprocess	wall 120 s / idle 60 s	COMMIT FAILED at 1–2 min if pre-commit hook silent
git push subprocess	wall 230 s / idle 90 s	PUSH FAILED at 1.5 min if pre-push hook silent
CreatePR gadget	240 s outer	Harness-invoked tool died at 4 min regardless

A recent agent session lost 10+ minutes looping on the 120 s push timeout when lefthook ran the full test suite. This PR lifts all three.

Scope (explicit)

In scope, per approved plan:

• src/gadgets/github/core/createPR.ts — drop 4 timeout constants, pass 0/0 on both runCommand calls
• src/gadgets/github/definitions.ts — set outer timeoutMs: 0

Out of scope (deliberately not touched):

• runCommand DEFAULT_IDLE_TIMEOUT_MS / DEFAULT_WALL_TIMEOUT_MS — other callers (agent worker repo fetch/checkout/reset, setup.sh, npm install, tsc) still want defaults to catch genuine hangs
• Every other timeout in the inventory: HTTP clients, watchdogs, VerifyChanges, diagnostics tsc/biome, tmux, Sleep, router workerTimeoutMs. Each guards a real hang case.

Test coverage

Added:

• 2 tests in `tests/unit/utils/repo.test.ts` pinning that `wallTimeoutMs: 0` and `idleTimeoutMs: 0` actually disable the kill path (10 min of silence / 10 min of continuous output → no SIGTERM). Previously the disabled-path was undocumented by tests; only the enabled path was covered.
• 3 regression tests in `tests/unit/gadgets/github/core/createPR.test.ts` pinning that `git commit` + `git push` pass `wallTimeoutMs: 0, idleTimeoutMs: 0`. A future author who "just adds a safety cap" hits a red test with context.

Updated:

• `definitions.test.ts` "every definition has a timeoutMs > 0" → "every definition has a non-negative timeoutMs (0 is the disabled sentinel)", with rationale for when 0 is appropriate.
• `definitions.test.ts` "CreatePR has a 4-minute timeout" → "CreatePR has timeoutMs disabled (0) — hooks may run for minutes".
• Deleted the three outdated spec-013 contract tests that asserted the old per-caller timeouts were finite. Kept the captured-output assertions from that block — those still pin that pre-commit / pre-push hook output is surfaced.

Totals: 6400 unit tests pass (up from 6398), typecheck clean, lint clean (0 warnings).

Test plan

• `npm run typecheck` — clean
• `npm run lint` — clean, 0 warnings
• `npx vitest run --project unit-core --project unit-api` — 6400 pass, 23 skipped
• Manual smoke: rig `lefthook.yml` pre-push to `sleep 180`, run `cascade-tools scm create-pr`, observe heartbeat continues past 90 s / 230 s and push completes cleanly. (Owner can verify on satellite after merge.)

Rollback

Single commit. Revert restores the four subprocess timeout constants and the gadget `timeoutMs: 240000`.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!