rubyspec-gate's expected-PASS list can be split into shards

[yosefbennywidyo]

make gate takes a while even in parallel: splitting it across fork Actions jobs (corpus in TEST_SHARD slices, one job per rubyspec-gate suite, props, bench, optcarrot) gets it down to about 11-14 minutes, except for one leg. rubyspec-gate's language suite has no way to split: it runs every expected-PASS example (1,025 today) in one job, and stays the longest leg even with 4 workers.

RUBYSPEC_SHARD=k/n keeps only every n-th expected-PASS example per suite (0-indexed offset k-1), so a CI fork can run language (or any suite) as two or more parallel jobs instead of one. Each shard still runs the gate's own checks per suite ("every listed example ran", "no regression"); running every k from 1 to n covers the full expected-PASS list exactly once, so the shards' combined result is the same gate the unsharded target runs. Without the variable, make rubyspec-gate is unchanged.

The variable is validated before use: CodeRabbit caught, on the fork PR, that an out-of-range RUBYSPEC_SHARD (k=0, k>n, or n=0) made the awk filter match no line, and a non-numeric one crashed awk outright -- in both cases the recipe kept going with an empty expected-PASS list, and the gate's own "ran == want, no regression" check passed vacuously ("all 0 expected-PASS examples still pass" reads as a real pass). The variable is now rejected up front unless it is k/n with both positive integers and k <= n.

Verified locally on core/range (200 expected-PASS examples): a 1/2+2/2 split and a 1/3+2/3+3/3 split each reproduce the full example set with no overlap and no gaps, every shard passes with 0 regressions, and a run without RUBYSPEC_SHARD is byte-for-byte identical to the one before this change. The same partition math checked against language's actual expectations file (1,025 -> 513 + 512, union equal to the full list). The validation was checked against the actual failure modes (RUBYSPEC_SHARD=bogus, =0/2, =5/2) before and after the fix.

Related: #6762 proposes widening what the gate covers; no overlap with this change, since #6762 is about which examples get enrolled, not about splitting execution of an already-enrolled suite.

Gate ran sharded on the fork: https://github.com/yosefbennywidyo/spinel/actions/runs/37731651385 (17 jobs, all green, ~14 min). The commit's own Gate: trailer records it:

Gate: green tree 196d31629368 master 9c7ea3c (linux-x86_64 gcc-13.3.0) tests 6421/0

---

🤖 Generated with Claude Code

Summary by CodeRabbit

• New Features
  • Added support for splitting expected passing tests into numbered shards. Shards cover the full test list when run together, while each shard retains its existing run-count and regression checks.
  • Invalid shard values are rejected; leaving the setting unset runs the full test list.

[coderabbitai]

Note

Currently processing new changes in this PR. This may take a few minutes, please wait...

⚙️ Run configuration

• Configuration used: Organization UI
• Review profile: CHILL
• Plan: Advanced
• Run ID: bf39e649-5db0-402f-9d9a-fc79924bc3e0

📥 Commits

Reviewing files that changed from the base of the PR and between 106c965 and 819e369.

📒 Files selected for processing (1)

• Makefile

 _____________________________________________________________________________________________________________________
< Use a project glossary. Create and maintain a single source of all the specific terms and vocabulary for a project. >
 ---------------------------------------------------------------------------------------------------------------------
  \
   \   (\__/)
       (•ㅅ•)
       / 　 づ

• Autopilot · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}