CI lanes and Bazel caches¶
Ordinary Linux and macOS CI retain inexpensive RF unit coverage and one native inverse-design
smoke. Solver sweeps, gradient checks, multi-iteration optimizations and coupon fitting run in
RF nightly, at 03:17 UTC daily on ubuntu-24.04-arm and macos-latest, or through Run workflow.
Scheduled runs use main; manual dispatch can select a branch. Before the new workflow is
registered on the default branch, use CI → Run workflow → rf_nightly (or
gh workflow run ci.yaml --ref BRANCH -f rf_nightly=true); the existing CI workflow calls the
same RF lane only for that explicit manual input. This workflow does not trigger
on pull requests or pushes. Both matrix jobs report failures independently.
flowchart LR
PR[Pull request / main push] --> Ordinary[Linux + macOS ordinary tests]
Ordinary --> Smoke[One native forward / adjoint / MMA smoke]
Timer[Nightly / manual dispatch] --> RF[Linux + macOS expensive RF lane]
Restore[Newest compatible cache snapshot] --> Ordinary
Restore --> RF
Ordinary --> Save[Save completed outputs, even after test failure]
RF --> Save
Running the lanes¶
bazel test //... --config=ci
bazel test //tests/unit/rf/... //tests/unit/rf_coupons/... //tests/e2e/rf/... --config=ci --config=rf-nightly
On a shared workstation, use nice -n 10 and --config=lowmem as described in
DEVELOPERS.md. Default and quick configurations exclude rf-nightly;
--config=all includes it, while Bazel’s existing manual gate still applies. Explicit manual
labels remain runnable. Native requirements, CPU reservations, sharding and per-test timeouts
are preserved. No solver or production behavior changes.
The ordinary test_inverse_smoke makes a real native forward/adjoint evaluation and an accepted
MMA update on the tiny two-port problem. It asserts finite results, a changed design and exact
repeatability across two fresh calibrations. Missing native code fails the Bazel target.
Longer convergence, resume, export, gradient and multi-backend checks remain in nightly coverage.
Complete inventory of affected packages¶
The source of truth is tools/bazel/rf_lanes.bzl. Each new globbed RF test must be classified;
each CI test job also compares Bazel’s expanded rules with this inventory, including manual
variants. A missing target, an unclassified addition or an incorrect gate fails that check.
Cheap RF order-0 and Palace unit packages outside this inventory retain their existing selection.
//tests/unit/rf¶
Lane |
Targets |
|---|---|
ordinary |
|
nightly |
|
manual |
|
kicad |
|
//tests/unit/rf_coupons¶
Lane |
Targets |
|---|---|
ordinary |
|
nightly |
|
manual |
|
kicad |
|
//tests/e2e/rf¶
Lane |
Targets |
|---|---|
nightly |
|
manual |
|
The five full RF e2e cases, four NumPy gradient references, double-resolution microstrip and coupon cross-section test were already manual; they stay explicitly manual because they need larger budgets or additional tooling. KiCad tests keep their dedicated gate. Nothing is removed.
Evidence and limits¶
The observed pre-split macOS run at commit 84b459979fbf6d7804f5db03b1aedd34b4e6743a
reported 44m17s in its ordinary Test step, versus 8m48s on Linux. Mac executed 291 tests and
reused two cached test results; Linux executed six and reused 287. Several RF targets reported
one to three minutes; coupon test_fit reported 340.9s and coupon test_order0 176.0s.
These are individual durations from a shared CI run, not additive wall-time savings or a
controlled benchmark. Fresh hosted runs must establish the new lane’s runtime and cache hits.
The old macOS fallback cache came from September 30. Compiler/image drift was not established: the old producer and observed consumer reported the same hosted image version and Bazel 7.7.1. Source, BUILD and dependency changes can legitimately invalidate actions. Cache restoration also does not eliminate Bazel analysis or runfiles reconstruction.
Cache lifecycle and trust¶
The ordinary test jobs and RF lane use explicit actions/cache/restore@v6 and
actions/cache/save@v6 composite actions. Download caches keep a stable key; a missing exact
key is saved after a completed test invocation, including a failure. Download restore also
accepts the previous per-platform namespace. Build snapshots include
OS, architecture, hosted image version, compiler/SDK fingerprint and hashes of Bazel/configuration/lock
inputs. Each save adds the run ID, attempt and job, avoiding an immutable key freezing the first
snapshot. Restore selects the newest matching compatibility prefix. Bazel still checks each
action digest; no full output base, sandbox or runfiles tree is cached. Hosted image updates
start a new build-cache partition while keeping repository downloads reusable.
A completed failed test still fails its job. Only cache uploads tolerate upload errors, and
cancelled jobs or tests that never started do not save. Bazel’s default --cache_test_results=auto
reruns failed tests. No timeout is raised and no test uses continue-on-error.
Fork pull requests restore only. Same-repository pull requests may save under GitHub’s isolated
PR merge-ref scope; trusted main push, scheduled and manual runs populate caches available to
future PRs. No pull_request_target, credentials or wider permissions are introduced. Unique
build snapshots use normal GitHub cache retention/eviction; concurrent producers may save
partial snapshots, and no cache hit or runtime improvement is guaranteed.