refactor(ci): replace inline fixture-collector bash with a typed, tested script

[JPHutchins]

What

Replaces the ~90-line inline bash in the Collect fixtures step with a single-file, dependency-free Python script (.github/scripts/collect_fixtures.py) run by uv (PEP 723), plus a new check-script CI job that gates it on ruff, ruff format --check, mypy --strict, and pytest (36 tests). The build job now needs: check-script.

Why

The step was untyped, untested string-munging (find + sed + grep + jq). The rewrite is a pure functional core (parsing, transport/group classification, canonical naming, manifest shaping) with a thin IO shell (listing, copies, manifest write), and an exhaustively-matched tagged-union artifact model (Exe | MergedHex | Hex | Elf).

Behaviour

Manifest output is unchanged for all real Twister output — verified by an end-to-end uv run against a synthetic twister-out (the emitted qemu_cmd is byte-identical to the old jq form) and by the unit/integration tests. Latent bash bugs fixed in passing, none observable on real inputs:

• merged.hex silently overriding a co-present .exe
• run pointing at whichever path last set $artifact instead of the executable
• the QEMU load flag chosen by sniffing the filename suffix rather than the image kind
• a build-dir name without exactly one dot silently mis-sliced into config/target
• silently taking the first of several merged_*.hex

The first two are mutually-exclusive in practice (native_sim vs sysbuild), so the change is behaviour-preserving; the rewrite just makes them structurally impossible.

Also — out-of-date action pins

• astral-sh/setup-uv@v8 (new)
• softprops/action-gh-release@v2 → @v3 (Node 20 → 24 runtime bump; safe on ubuntu-24.04)
• checkout / setup-python / upload-artifact / download-artifact are already at their current majors — left as-is.

Test plan

• ruff check and ruff format --check clean
• mypy --strict clean (script + tests)
• pytest 36/36 (incl. doctests + an end-to-end main())
• uv run end-to-end manifest matches the old jq output
• CI: check-script + both build legs (native_sim, arm) green on this PR

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR refactors the CI “Collect fixtures” logic by replacing a large inline bash script in the build workflow with a typed, dependency-free Python script (PEP 723, run via uv), and adds a dedicated CI job to lint, type-check, and test that script before the build runs.

Changes:

• Add .github/scripts/collect_fixtures.py to collect Twister outputs into canonically named fixtures and emit a sorted JSON manifest.
• Add .github/scripts/test_collect_fixtures.py plus .github/scripts/pyproject.toml to enforce ruff, mypy --strict, and pytest coverage for the script.
• Update .github/workflows/build.yaml to introduce a check-script gating job, run the collector via uv, and bump the release action pin.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
.github/workflows/build.yaml	Adds check-script, gates build on it, switches fixture collection to uv run ...collect_fixtures.py, and updates the release action pin.
.github/scripts/collect_fixtures.py	New typed fixture collector script implementing naming, artifact selection, manifest shaping, and IO.
.github/scripts/test_collect_fixtures.py	New unit/integration tests defining the behavioral contract for the collector.
.github/scripts/pyproject.toml	Tooling config for ruff/mypy/pytest for the script + tests.
.github/scripts/.gitignore	Ignores local Python tooling caches for the scripts directory.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[JPHutchins]

Thanks — addressed both in f7c75c7:

1. Python pin in check-script — good catch. The gate ran uvx pytest on the runner's default interpreter (3.12 on ubuntu-24.04), not the 3.13 the script targets. Set UV_PYTHON: '3.13' on the job so ruff/mypy/pytest all run under the same runtime as the build leg.

2. .exe vs merged.hex — fair, the wording overstated it. I narrowed the fix to the real contradiction: select_artifact now raises when a dir has both a native_sim .exe and a merged.hex (host vs flashed-MCU target classes that never coexist). I deliberately did not raise on "multiple image kinds" generally, because real sysbuild dirs legitimately carry merged.hex + zephyr.hex + zephyr.elf together — those still resolve via the merged > exe > hex > elf ladder. Added a regression test for each case.