CI runs ~200 times a day at ~47 runner-minutes each (Blacksmith, sponsored).
- New `changes` job (scripts/ci/changes.sh). code=false only for a pull_request or
merge_group whose diff touches nothing but prose no build/test/difftest reads
(plan/findings/, plan/upstream-issues/, plan/*.md, site/, docs/, assets/, top-level
*.md). Those skip clippy/test/wasm/publish-dry-run/deny and the difftest jobs; the
required checks report "skipped" (passing). fmt (with the frozen-archive guard) and
ci-scripts always run. Any doubt (no base, fetch failure) means code=true.
- Pushes to main (full=false) were already tested on the identical tree by the merge
queue: they now run only what publishes main's artifacts, difftest-sweep (the
difftest-reports the next PRs' exact-base ratchet reads) and the solids report step.
- `difftest` aggregate: passes on non-full runs when the sweep succeeded or was skipped;
on full runs it still requires every part.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QaDuBSC1mqS9DnQoEsr7LJ
Keeps #983's known_mismatches.toml guard next to the archive guards in
check-frozen-archives.sh, and both sides' wording in AGENTS.md and
plan/findings/README.md. split_archives.py leaves the issues archive's own
"Frozen archive" notice (3 lines, dead relative link) out of the README; the
check lists those lines and compares the rest.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QaDuBSC1mqS9DnQoEsr7LJ
scripts/split_archives.py (run with python3 -I) splits the frozen plan/findings.md
(425 `## ` sections) and plan/upstream-issues.md (436 `#### ` entries) into one file
per record under plan/findings/ and plan/upstream-issues/, moves the archives' intro
and "Note: duplicate IDs" text into the READMEs, and deletes the archives.
`split_archives.py check` proves the split lossless against the archives at 4fe6f4d51.
site/build.py reads issues only from the directory (intro from its README);
check-frozen-archives.sh now fails if either archive comes back.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QaDuBSC1mqS9DnQoEsr7LJ
Known mismatches move from the shared crates/difftest/known_mismatches.toml
to crates/difftest/known_mismatches/<unit id>.toml (same [[mismatch]]
schema, the 24 entries copied verbatim). for_unit reads only that unit's
file (missing = no entries) and rejects an entry whose `unit` differs from
the file name. A test checks every file names a plan unit and is non-empty,
and check-frozen-archives.sh fails if the old shared file comes back.
SolidCraft cases were already one JSON file per issue; the conflicts came
from a hard-coded total case count in every_case_file_loads that each case
PR bumped. The test now derives the total from the files.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QaDuBSC1mqS9DnQoEsr7LJ
scripts/ci/check-frozen-archives.sh runs in the required fmt job and fails a
change that adds lines to plan/findings.md or plan/upstream-issues.md once the
base has the freeze notice (deletions pass). Updates to an earlier record,
including archive entries, are new addendum files that link to it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QaDuBSC1mqS9DnQoEsr7LJ
Split the serial difftest job: the exact-base artifact is fetched and validated in
one job while the green sweep and the mandatory perturbation check run in their own
jobs. A new difftest --compare-reports mode applies the same ratchet as --baseline to
the sweep's finished reports, and the required check name 'difftest' becomes an
aggregator that fails unless every part succeeded.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011gJAwcLRBPQXYLmvXd3pWf
Resolve speculative queue bases before selecting the exact trusted push/main artifact, and allow 25 minutes for pending main CI. Keep ancestry provenance and add queue, retention, integrity and timeout regressions.
Co-Authored-By: OpenAI Codex <noreply@openai.com>
A known mismatch may remain known or become a real pass, never deferred or rejected. Retained cases keep numerical value coverage. Legacy indexed integer mismatch evidence proves presence without fabricating omitted comparison counts; ambiguous known integer coverage fails closed.
Co-authored-by: Codex <codex@openai.com>
Preserve the original regression-gate author commits while retaining current CLI, report, known-mismatch and CI behavior. Validate trusted exact-base artifacts before applying case retention and the unchanged value-share floor.
Co-authored-by: Codex <codex@openai.com>
Pushes to main used the org's self-hosted runners (two of four are offline) while every PR
already ran on Blacksmith's sponsored runners. Move main onto Blacksmith too, give
difftest 16 vCPU, and remove the self-hosted-only steps and warm-target.sh. Caching is
actions/cache throughout. Fallback when the sponsorship ends: blacksmith-* labels back
to ubuntu-latest.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Pfe7dodPG5rpuN9B8QcnuX
Review (#215): a failed fetch or extraction of the base commit's reports only
warned, so --green then ran without --baseline and the ratchet was silently
off. Move the step into scripts/ci/difftest-baseline.sh: a pull request must
have a base; for a real base, fetch, extraction and "has reports" must all
succeed or the job fails. Only a push without a base commit (all-zero
`before`) or a manual run skips the ratchet, with a notice.
scripts/ci/test-difftest-baseline.sh exercises it in a throwaway repo (real
base, failed fetch for PR and push, base without reports, missing/zero base,
bootstrap cases), run by a new ci-scripts job. Restoring the old warn-only
behaviour makes it fail.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* ci: run trusted jobs on the self-hosted runners
Pushes to main and pull requests from branches of this repository now run fmt,
clippy, tests, wasm, the publish dry run and difftest on the org's self-hosted
runners, which keep their Cargo build directory and C++ reference build between
jobs. cargo-deny (a container action), fork pull requests and Run workflow stay
on GitHub's runners. The C++ cache on GitHub is no longer shared with the
self-hosted build, whose newer compiler and glibc must not reach GitHub's runners.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* ci: re-run on the self-hosted runners
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
bherbruck