build: cut the gates job's run time

[bpowers]

Why

The gates job is the longest job in CI: about 20 minutes on main (5.5 to 8 to compile, 10 to 13 to run on the last four runs), past the macOS jobs at about 15.

What

• The budget property test draws 40 generated models a run, not 200. On a runner this one test was the run step: its cases run one after another on one thread, about 11.5 minutes, and for the last 6 of them it was the only test still running. Each run draws new shapes and replays every saved failing seed, so coverage accumulates across runs.
• scripts/gates.sh with no arguments runs the ignored tests only (47 today), not the whole engine suite again. It still builds with the ext_data feature (the Excel data provider) and runs that feature's own tests by name (data_provider::, 43 tests, under a second): they are not ignored, and no other job compiles them. scripts/gates.sh <filter> is unchanged.

Evidence

Measured on this PR's own gates job:

	Compile	Run
main (last four runs)	5.5 to 8.3 min	7.3 to 12.7 min
ignored tests only	8.1 min	14.0 min
plus 40 cases	6.9 min	6.3 min

Running only the ignored tests saved nothing by itself: the already-run tests cost seconds once optimized. The property test was the cost. After the change the unit gates finish in 4.6 minutes and the integration gates in 1.7.

What this does not establish

• The compile step, now the larger half, is unchanged.
• 40 cases a run finds a rare overrun later than 200 would. The two this gate found this week were each hit within a few dozen cases.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LYR7N9WZwQuFnZFW8JDUgx

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-04T06:57:13.203397Z	22c2ae0	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[claude]

Review

No blocking issues found. The script correctly:

• Preserves the single-invocation path for user-supplied filters (--include-ignored "$@").
• Runs ignored tests across the engine, then the data_provider:: tests separately, so a #[ignore]d test anywhere in the crate (including hypothetical future ignored data_provider tests) still runs via the first pass.
• Keeps the second run going even if the first fails (|| status=$?), and exits non-zero if either failed.
• Compiles both passes under one --no-run build, since the targets and features are identical.

Verified that only data_provider/ contains #[cfg(feature = "ext_data")]-gated code (the single hit in json.rs:4175 is a user-facing string label, not test code), so the data_provider:: filter catches every ext_data-only test.

Overall correctness verdict

Correct.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 92.90%. Comparing base (e707b69) to head (5f0d7a7).

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #1086   +/-   ##
=======================================
  Coverage   92.90%   92.90%           
=======================================
  Files         313      313           
  Lines      194144   194144           
=======================================
+ Hits       180363   180369    +6     
+ Misses      13781    13775    -6     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

[claude]

Review

Reviewed the trim to scripts/gates.sh and the proptest cap drop. No blocking issues found.

What I verified:

• data_provider:: as a libtest substring filter matches all ext_data-sensitive tests. The three files that reference feature = "ext_data" are src/data_provider/csv_provider.rs, src/data_provider/mod.rs, and tests/integration/mdl_vensim_truth.rs; the first two are covered by the filter and the third's only cfg!(feature = "ext_data") branch is inside the #[ignore]d corpus gate (the_mdl_reader_matches_vensim_over_the_corpus), which still runs via --ignored. The non-ignored companion each_class_is_told_from_the_others only classifies four paths, none in READS_A_SPREADSHEET, so it is insensitive to the feature flag either way.
• Exit-code handling: run --ignored || status=$? followed by run "${FEATURE_TESTS[@]}" || status=$? and exit "$status" correctly surfaces a non-zero exit if either run failed, without one run masking the other, matching the stated intent.
• Filter branch (scripts/gates.sh <filter>) still passes --include-ignored so users can name an ignored test by name — behavior preserved.
• Proptest 200 → 40: the ignored-reason string is updated to match ("over 40 generated models"), and the comment's claim that per-run coverage grows across runs via the saved failing seeds matches how proptest handles the regressions file.

Overall correctness verdict: correct. Existing behavior is preserved for the filter and --no-run paths, no coverage is dropped relative to what any other CI lane compiles, and the proptest cap change is intentional with a clear rationale.