Skip to content

ci: the gate-speed changes together (#470, #476, #477, #478, #467, #479) - #483

Merged
humuhumu33 merged 20 commits into
release/v0.4from
ci/gate-speed
Nov 12, 2026
Merged

humuhumu33 merged 20 commits into
release/v0.4from
ci/gate-speed

Conversation

@humuhumu33

Copy link
Copy Markdown
Contributor

Lands six CI and conformance changes together, with one full gate instead of six.

PR Change Effect measured on its own
#470 The BS evaluator differential runs on every processor BS-15 145 → 100 min; whole run 149 → 134 min (run 43046966603)
#476 The bootstrap project is exported once per run and handed to every part, with a custody record Removes one bootstrap verification (1,173–2,205 s) from each of 19 jobs (dispatch 43127877709, all green)
#477 The Windows host runs as five lanes plus a join The 106–126 min Windows job, the longest after #470, is split into lanes
#478 cargo xtask check-freshness in honesty A tree with stale generated files fails in minutes, before the fan-out (draft run 43142198451)
#467 regenerate.yml: a slice's whole landing on runners Regenerated #452 and the BS-25..34 batch (#480, run 43112441927)
#479 The five duplicate keys of .github/bdd-parts.json A strict parse equals the old parse

How they were combined

Verified locally on this tree

  • Rust: cargo fmt --check, workspace clippy with -D warnings, validate-model and validate-spec-links (493 IDs), and the repo-conformance library tests (44) pass.
  • Conformance: the bdd meta-gate target (5) and all 17 conformance_rp_* cases pass, RP-05, RP-13 and RP-14 among them.
  • Freshness: cargo xtask check-freshness passes (23 build oracles, 272 inputs, 164 export packages).
  • Python: both unit-test suites in .github/scripts pass, Intel factory and host lanes.
  • Plans: bdd-parts.json has 37 parts, no duplicates, and every BS-25..34 part.

The full gate on this PR measures the six together. Merging it lands #467, #470 and #476 to #479.

🤖 Generated with Claude Code

humuhumu33 and others added 20 commits November 11, 2026 13:16
A dispatch on a ref runs the production export, the verification records and
the other oracle writers as three jobs from clean checkouts of that commit, and
a final job combines their patches into one, refusing two that write one path.
It reads the repository and writes nothing to it; the patch is reviewed and
committed like any other written output (§28.3).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
GitHub dispatches only a workflow its default branch carries, so a push to a
branch under regenerate/ starts it from any commit; regenerate/atlas/ includes
UOR Atlas's records.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Each bdd-bootstrap-* part of the gate is a single case, and its evaluator
differential ran its corpus on one thread of a four-processor runner. The
runs are independent, so the calling thread and helpers from one budget the
process shares each take the next run. The verdict and the disagreements
reported are the sequential ones; PD-08's stack measurement keeps every run
on the measured thread.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…bounds

A fourth job, after combine, applies the regenerated tree, runs the parts
that hold a crate::counts! check under LEXLEAN_COUNTS=fast, writes the moved
counts (cargo xtask counts --write) and the model documents, and measures the
stack bound of each recursive package the record lacks with a whole run of
the case that names its root. The patch it returns is the whole landing.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Every bdd part that runs a bootstrap case exported the bootstrap project
itself, a whole verification of 1,173-2,205 s per part (19 jobs, about
569 runner-minutes per gate in run 43046966603). The new `bootstrap export`
job runs it once after honesty at the examples leg's profile (`cargo xtask
export-bootstrap`), requires the committed packages, and uploads the bundle
with a custody record binding repository, run, attempt, commit, tree, pins,
toolchains and the bundle digest. Each leg starts a fetcher that takes the
artifact when a case first needs it; `exported("bootstrap")` checks the
record against its own run and checkout and replays the bundle, and exports
as before when LEXLEAN_BOOTSTRAP_EXPORT is unset. RP-14 checks the job, the
fetcher and the record against planted changes.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Windows job ran the whole workspace in one `cargo test` and took
106-126 minutes, the longest job of the gate; 115 minutes of it were the
conformance harness (482 cases, 6,924 s). It now runs as five jobs on
windows-2022: the four conformance lanes of .github/intel-host-lanes.json
(the partition honesty already proves complete) and a producer lane with
every other test and doctest. Each lane builds from its own checkout and
records the host's listing and what it executed; the join on ubuntu keeps
the context `build and non-Lean tests (x86_64-pc-windows-msvc)` and passes
only when the lanes are this run's and commit's, disjoint, and together
equal the host's whole listing. RP-14 checks the lanes and the join against
plants; test_host_lanes.py runs the join's refusals in honesty.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ss`, SPEC.md §9.4)

Of 248 ci runs this month 84 failed, many of them full gates on trees
whose generated files were not regenerated; they failed hours in. The new
task checks, without Lean and in about 3 s, that every input a committed
build oracle (expected/build/manifest.json) records still has its digest
and length, that every such project's lock passes `lexlean lock --check`,
that every oracle and committed export package names this compiler's
semantics, and that every committed export package names its project's
committed semantic ID. Honesty runs it before its recipes, so a stale
ready pull request fails before the macOS and Linux fan-out starts. It
replaces no check of the gate.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
bootstrap-runtime, preservation-lists, preservation-runtime,
preservation-support and preservation-prims each appeared twice with
identical values; a JSON parse merged them silently. The parsed plan is
unchanged.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
# Conflicts:
#	CHANGELOG.md
#	xtask/Cargo.toml
# Conflicts:
#	CHANGELOG.md
#	SPEC.md
#	xtask/Cargo.toml
#	xtask/src/main.rs
# Conflicts:
#	.github/workflows/ci.yml
#	CHANGELOG.md
#	crates/conformance/src/cases/repository.rs

@afflom afflom left a comment

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Bounded integration review at f3b25f2, based on actual release aaf8455. The 16-path composition retains both custody and Windows-lane checks, freshness ordering, the original gate commands and all 37 plan parts. Reusing the individual reviews found no additional composition-only blocker. Four previously reported P2 findings are still present:

  1. export_custody.rs:381–393: take reaches identity/tool-version checks and verify_artifact without the existing support::env_shared guard. The production_export shortcut bypasses the guarded export path. Acquire the shared guard after waiting and the environment-independent digest, and retain it through identity/replay; avoid putting the long wait or recursive fallback initializer under it. This is a source-reachable race with the existing exclusive PATH-sentinel tests, not a claim that this CI run observed it.
  2. bootstrap.rs:1035–1094: helpers share N−1 permits but each of C concurrent callers also evaluates uncounted work, permitting C+N−1 evaluators (N=4,C=2 permits5). Account for callers, or explicitly narrow the claimed helper-only budget and its intended local behavior. Single-case hosted parts do not establish the advertised process-wide bound.
  3. regenerate.yml:123–132: combine still applies patches without intersecting their .files manifests first. The retained collision demonstration already shows disjoint hunks in one shared path can both apply. Reject overlapping writer path sets before application; reuse that evidence and preserve binary patches.
  4. VERIFICATION.md is unchanged. Record the already-performed direct check-freshness plants, commands, actual refusals/restoration and scope. No new plant execution or unrelated RP-14 plant is requested.

Please batch these narrow owner repairs before final-head acceptance and landing. The individual old-head gates remain valid only at their recorded scope; current full run43198040363 is still active and must not be aggregated with them or canceled partial runs. The PR's individual timings are not yet a measured combined-tree speedup.

Receipt /workspace/work/pr483-bounded-integration-review/receipt.json records the exact source bindings and limits. No competing patch, producer, test, gate, log download, rerun or cancellation was performed. Existing Claude ownership of implementation and qualification is preserved.

@afflom afflom mentioned this pull request Nov 12, 2026
@humuhumu33
humuhumu33 merged commit 465b8ab into release/v0.4 Nov 12, 2026
71 checks passed

afflom commented Nov 12, 2026

Copy link
Copy Markdown
Member

Actual landing readback: 465b8ab has ordered parents [aaf8455, f3b25f2] and tree bd8e6423d8b58e1163a64f5617dd08fe88ff3daf, equal to the reviewed head. PR run43198040363 passed all70 jobs and772 steps in attempt1, including same-run honesty; repro43198040382 passed. Post-merge push runs are separate evidence. GitHub reports all six component PRs merged, not superseded-unmerged.

The four source findings in review5872966303 were unresolved before this external landing. They are now retained in #489 with concrete bounded next steps for the existing batch owner; passing CI is not a source-clearance claim. No completed gate or demonstration was repeated. #269/#274 and other unfinished release parents remain open.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants