[loom-fleet-dispatch]

Summary

klt lvs on very large flat compares could run for tens of minutes with no output. This adds an opt-in deadline (options.compare_timeout_s) that yields a structured inconclusive report (exit 4), plus a stderr heartbeat. Native hierarchical compare and ambiguity stats stay out of scope, as the issue specified.

Changes

• lvs_timeout.py (new): option parsing, forked-child deadline runner, heartbeat.
• lvs.py: _compare_netlists takes a timeout and context; on expiry the report is status: "inconclusive" with one compare.timeout entry (details.code: "compare_timeout", elapsed time, device/net/pin counts). The option is echoed and replayed by --rerun only when set.
• Docs ("Large flat designs" section with the manual per-cell recipe, option row), response schema, CHANGELOG, import-boundaries entry.

Acceptance Criteria Verification

• Timeout gives inconclusive, compare_timeout diagnostic, exit 4, no leftover child: test_compare_timeout_reports_inconclusive_issue_4165 (asserts multiprocessing.active_children() == []).
• Option omitted leaves output unchanged; value not reached matches the untimed run: test_compare_timeout_not_reached_matches_untimed_issue_4165.
• Invalid values (0, -1, string, bool, NaN) are request errors: test_compare_timeout_invalid_value_rejected_issue_4165.
• Heartbeat goes to stderr only; stdout is valid JSON: test_compare_heartbeat_goes_to_stderr_only_issue_4165.

Test Plan

TDD: tests written alongside the implementation, not before it. Ran the new tests and the existing inconclusive/rerun/internal-error tests in tests/test_lvs.py, plus ruff, check_complexity_baseline.py and check_import_boundaries.py. Full suite left to CI.

Live verification NOT performed. The ~50k-device repro was not run; tests use a patched slow compare. Known trade-offs:

• A compare that finishes inside the budget is repeated in-process, because the comparer's logger holds native objects that can't cross a process boundary. A run near the limit can take up to 2x its compare time, which is why the option is opt-in.
• Without the timeout option the heartbeat comes from a thread and is best effort. It depends on the native compare releasing the GIL, which I did not verify.

Closes #4165

🤖 Generated with Claude Code

loom dashboard

[loom-fleet-dispatch]

PR #4166's loom:reviewing claim is held by the judge of sweep step-judge-4166-1792971733 on host host-d0749b00. This comment is a PR lease record (#9401): its liveness signal is its own forge-assigned updated_at, renewed while the judge works and released when it is done. Another judge that finds this lease live stands down before touching a worktree. See defaults/docs/lease-record.md § "PR-phase leases".

[loom-fleet-dispatch]

✅ Approved!

CI is green on this head (forge wait-checks: GREEN); merge state CLEAN; no formal reviews or inline threads outstanding (check-review-feedback.sh: CLEAR).

Reviewed

• lvs_timeout.py: option parsing rejects 0/negative/NaN/inf/bool/string and non-klayout engines as request errors. The forked child is terminated, then killed, then joined on every exit path. Child failure or EOF becomes an LvsError, never a verdict.
• lvs.py: _guarded_compare turns a timeout into a single compare.timeout mismatch and status inconclusive. In the parameter_tolerance path the logger is empty after a timeout, so no snaps are collected and the loop exits without a second compare. The option is echoed and replayed by --rerun only when set, so output without it is unchanged.
• Docs, response schema, CHANGELOG and the import-boundaries entry are consistent with the code.

Test execution

• pytest tests/test_lvs.py -k "4165 or 4134" → 11 passed (run locally in the builder worktree).
• Full suite not run locally; relying on green CI.

Not verified (as the PR itself discloses)

• The ~50k-device repro was not run live. The tests patch a slow compare, so the 2x cost of a compare that finishes inside the budget and the thread-heartbeat GIL behaviour of the native compare are untested. Neither blocks: the option is opt-in, the heartbeat is documented best-effort, and the fork-based timeout path does not depend on the GIL. This is not an issue-silent-omission shape, so the circular-fixture rule does not apply.

loom dashboard

[loom-fleet-dispatch]

PR #4166's (no label: the lease is the claim) claim is held by the champion of sweep step-champion_merge-4166-1792971853 on host host-d0749b00. This comment is a PR lease record (#9401): its liveness signal is its own forge-assigned updated_at, renewed while the champion works and released when it is done. Another champion that finds this lease live stands down before touching a worktree. See defaults/docs/lease-record.md § "PR-phase leases".

[loom-fleet-dispatch]

Champion: merging. All 6 safety criteria pass on head b3c6ebe: Judge approval fresh (verdict SHA matches head), merge-risk low (opt-in option, additive, tests cover timeout/invalid/stderr; no critical files), mergeable CLEAN, CI all green.

---

Automated by Champion role

loom dashboard