fix: kill child process tree on timeout + heartbeat step 4 (GAL-739)

[ADWilkinson]

Summary

• Group-signal the process tree in exec.ts so a claude -p grandchild actually dies when the step timeout fires (root cause of the GAL-738 80+ min wedges).
• Emit a 2-minute heartbeat log inside step 4 so operators can distinguish "slow" from "wedged" instead of staring at a frozen [4/8] Planning with Claude… line.
• Bump planMinutes default from 60 to 90 to cover deep-mode plans on monorepos with 3000+ line target files.
• New src/exec.test.ts asserts the grandchild-sleep scenario is cleaned up on Linux.

Why

Two recent oneshot --local --bg dispatches against personal/galleonlabs-zkp2p stalled inside step 4 for 96 min and 3h52m. The child claude -p process stayed alive because killProcessTree only signalled the top-level bash PID — the grandchild was reparented to PID 1 and kept burning CPU. Meanwhile, no log output was emitted after the step started, so the operator had no way to tell "still working" from "stuck". The 60-min default was also too tight for deep-mode plans that routinely run 40-80 min.

Relates to GAL-739 and GAL-738.

Changes

src/exec.ts — process group kill

• Spawn bash with detached: true so the child gets its own process group
• Signal both the leader PID and the negated PID (process group) in killProcessTree for both SIGTERM and the follow-up SIGKILL

src/steps/plan.ts — heartbeat

• Start a 2-minute setInterval before awaiting execOrThrow that logs elapsed time and the planMinutes limit via log.info
• Clear the interval in a finally block

src/config.ts — timeout bump

• DEFAULT_STEP_TIMEOUTS.planMinutes: 60 → 90

src/exec.test.ts (new)

• One test case: spawns bash -c "sleep 30" through the exec helper with a 2s timeout, waits 6s, then asserts via pgrep that no orphan sleep 30 process remains
• Skipped on non-Linux platforms

Test plan

Not run — typecheck and test execution deferred to CI / operator verification. The spec commands are:

bun run typecheck
bun run test
bun test src/exec.test.ts