Run Zephyr tests in their own workflow

[mikeysklar]

What

Zephyr tests move to their own workflow, out of the ports dependency chain.

Why

Board builds in every port waited on the Zephyr job, every PR.

# build.yml, before
ports:
  needs: [scheduler, mpy-cross, tests]
# tests = run-tests.yml = {run x4, zephyr}
# a called workflow is not complete until every job in it is

before:  tests{run x4, zephyr} ─> ports{espressif, raspberrypi,
                                        atmel-samd, nordic, zephyr-cp}

after:   tests{run x4} ────────> ports{espressif, raspberrypi,
         zephyr-tests                  atmel-samd, nordic, zephyr-cp}
         (gates nothing)

# ports still needs tests, so the saving is the delta between the zephyr
# job and the slowest run leg, not the zephyr job's full duration

Changes

File	Change
run-zephyr-tests.yml	New. Holds the zephyr job, body byte-identical.
run-tests.yml	Job removed. Now only run.
build.yml	Calls zephyr-tests, absent from ports' needs.
tools/ci_set_matrix.py	New workflow added to IGNORE_BOARD.

# ci_set_matrix.py: why the IGNORE_BOARD entry is not optional.
# Any .github file not listed there reaches this fall-through, so
# editing the new workflow would rebuild every board in the repo.
# run-tests.yml is already listed for exactly this reason.
            # Otherwise build it all
            boards_to_build = all_board_ids
            break

Needs a maintainer action

Branch protection lists required checks by name. This renames one.

Before	After
tests / zephyr	zephyr-tests / zephyr

# if the old name is required, the rename either blocks every PR on a
# check that never reports, or silently drops the requirement.
# I cannot see repo settings, so I cannot tell which.

Side effect of the rename

The rename also changes ci_changes_per_commit.py. Correct, but worth naming.

# tools/ci_changes_per_commit.py:181
    have_dependent_jobs = ["scheduler", "mpy-cross", "tests"]
...
                if any([name.startswith(job) for job in have_dependent_jobs]):
                    return {}

# `tests / zephyr` matched this prefix, so a failed Zephyr job disabled
# the resume-from-last-checked-commit optimization entirely.
# `zephyr-tests / zephyr` does not match. The result is discarded by
# ci_set_matrix.py, which reads only "ports" and "docs".
# Deliberately not added to have_dependent_jobs: nothing depends on the
# job, and it still runs on every PR, so no coverage is lost.

Why not gate the Zephyr job on changed files

It exercises core py/ and shared-*. Reordering is safe, skipping is not.

Trade-off

A red Zephyr job no longer stops board jobs from spending compute.

# the 4-way `run` matrix still gates ports. The Zephyr job is a weak
# canary for an espressif or atmel-samd build, and it cost every PR
# wall-clock time.

Hardware tested

None. Workflow change only.

How I tested it

pre-commit run --from-ref <base> --to-ref HEAD   # clean
python3 -c "import yaml, ..."                    # all three parse

# TODO: link this PR's run, board jobs starting while zephyr-tests runs
# TODO: zephyr and slowest run-leg durations, 3+ main runs, and the delta
# NOTE: this PR rebuilds every board, because it edits build.yml, which
#       is correctly not in IGNORE_BOARD. The new entry only affects
#       later edits to run-zephyr-tests.yml.

Scope

Job wiring only. The Zephyr job body is unchanged.

# two carried-over oddities left alone to keep the move byte-identical:
#   strategy: fail-fast: false   with no matrix
#   id: set-up-submodules        never referenced

Not in this PR:

• Caching the west workspace in deps/ports/zephyr-cp, which caches
  nothing today.
• Tiering the 29 Zephyr board builds: a subset on PRs, all on main.

AI assistance

Claude Code traced the dependency chain and wrote the diff.

[tannewt]

One comment we don't need. Good otherwise. This sort of thing is better documented by git history than inline code comment.

[tannewt]

Thank you!

[mikeysklar]

Here we go!