[eazene]

Checks that do real work were failing on the fixed 60-second check timeout, not on the work.

We run each executed check inside a container/VM launcher, so the 60 s also covers starting the sandbox and waiting
for a free slot. Multi-suite checks were taking about 54 s. One of them went over 60 s after a correct worker run:
the task failed, we re-checked it by hand, and the run was voided on our scoreboard. Workers already have a per-task
timeout_s; checks had no equivalent.

What this adds

An optional task field, check_timeout_s:

• Value. A JSON integer from 1 to 3600. When absent, CHECK_TIMEOUT_S (60) applies exactly as today, so existing
  manifests are unchanged.
• Validation. Anything else fails manifest loading with the task key: 0, negatives, 3601, 90.0, 2.5,
  "90", true or null. So lint, run, --baseline and --dry-run all refuse the manifest. A bool is
  rejected even though it is an int subclass.
• Use. Verifier._run_check uses the task's effective value; the process-group termination path is unchanged.
  The timeout message names the value (check timed out after <n>s).
• Visibility. The effective value is recorded in the run-state task record next to timeout_s and shown in the
  --dry-run listing.
• Dataclass. TaskSpec.check_timeout_s defaults to None, and the effective_check_timeout_s property reads the
  module constant at call time, so tests (and anyone patching the default) see the current value.

The README task-field table gains one row.

Tests

tests/test_check_timeout.py (9 tests, about 4 s), stdlib unittest, with RINGER_NO_SELF_UPDATE=1 for the CLI
cases:

• absent field → effective 60
• 1 and 3600 accepted
• every rejected value above, with the task key in the error
• with the default patched to 1 s, a task with 5 s passes a 2 s check, and a task without the field times out (the
  default is read at run time)
• check_timeout_s: 1 stops a 5 s check, and the message says 1s
• Manifest.from_obj refuses a bad task
• --dry-run shows check_timeout_s: 45 and check_timeout_s: 60
• lint exits non-zero on "90"

Full suite: 278 tests OK locally on macOS with Python 3.12.

Notes

• The cap. 3600 is there to catch typos such as an extra zero. Happy to change it, or drop it to match
  timeout_s, if you prefer.
• Real-world use. We've carried this as a local patch on a vendored copy and used it on real runs, with checks at
  120, 180 and 240 s. If it lands, we drop the patch.

🤖 Generated with Claude Code

https://claude.ai/code/session_01F74VKWWyJNxArUTPPaiBpb