fix(agent): enforce a foreground bash timeout (PRI-3251)

[ada-sen]

PRI-3251: Foreground bash tool timeout isn't enforced

Stacked on fix-pri-3243-bash-stdin-and-kill (base of this PR), which
made foreground bash spawns detached process-group leaders and closed
stdin, so an abort can kill a whole pipeline instead of orphaning it. This
PR reuses that same process-group kill for a new failure mode: a hung
foreground call with no abort in play at all.

The premise

The bash tool's own docstring has long claimed:

A sync command is subject to a runtime timeout (tens of seconds) and is
killed if it exceeds it

Nothing in the implementation enforced this. executeCommand only ever
killed the child process in response to context.signal's abort event —
there was no timer, no schema timeout/timeoutMs parameter, and no
host-level (agent loop / turn) timeout standing in for it either. A
command like sleep 999999 run in the foreground would simply hang until
something upstream (the caller aborting, or the whole turn timing out)
intervened — which for a bare tool call is never.

What changed

• Default timeout: 60s, matching the docstring's own "tens of
  seconds" language.
• Optional per-call override via a new timeoutMs schema field,
  bounded to [1_000, 600_000] ms (1s–10min). The upper bound matches the
  progressIntervalMs ceiling already established in this same file; the
  lower bound matches url_fetch's MIN_TIMEOUT and is low enough for a
  1s test to actually exercise the path.
• On expiry: SIGTERM the whole process group, then SIGKILL after a 2s
  grace period if it's still alive — the exact signal sequence and
  group-kill helper PRI-3243 added for abort, reused here via a shared
  scheduleGracefulKill(). A grandchild in a pipeline (e.g. sleep | cat)
  is killed too, not left orphaned.
• Result shape: the JSON result now carries timedOut: boolean and,
  when true, a timeoutMessage string ("Command timed out after Ns and
  was killed."). Partial stdout/stderr already captured before the kill
  are preserved unmodified in stdoutPreview/stderrPreview.
• Background bash calls are unaffected — they're routed to a job before
  ever reaching this code path, so the new timeout only applies to
  genuine foreground/sync calls.

Tests (packages/agent/src/tools/bash.test.ts)

Added a PRI-3251: foreground timeout is enforced block:

• a hung command (sleep 60) with timeoutMs: 1000 returns promptly with
  timedOut: true, a matching timeoutMessage, and the output produced
  before the sleep still present in stdoutPreview;
• a grandchild in a pipeline (sh -c '...; exec sleep 60' | cat) is
  confirmed gone (not orphaned) after the timeout fires;
• a fast command with the same timeoutMs: 1000 override returns well
  within budget with timedOut: false and no timeoutMessage.

Verified fail→pass on this branch: checking out the pre-fix
bash.ts from the fix-pri-3243-bash-stdin-and-kill base while keeping
the new test file reproduces exactly the bug this PR closes — the two new
timeout tests hang to their test-level timeout (they never see a
timedOut flag because nothing sets one), and the docstring-text
assertion fails because the docstring itself is unchanged. Restoring the
fix makes all cases pass.

packages/agent unit suite (excluding the integration/live-API test
files the repo's own test script separates out), lint, and typecheck
all pass on this branch.

Not in scope

• Background (background: true) bash jobs — these have their own job
  lifecycle and are explicitly excluded here.
• Any host-level (agent loop / turn) timeout — there isn't one today, and
  this PR only adds enforcement inside the bash tool itself.

[ada-sen]

Superseded — opened against upstream instead: obra#414 (stacked on obra#413). Branch left in place; not deleting it.