#273 #248/test the consumer install path and the shipped env in CI

[yunlishao-vibe]

Summary

Two CI gaps. Each has already let a broken release through with CI green.

#273: nothing ran the consumer install path

#238 changed the packaging, the installer wasn't updated to match, and the first to notice was a user running --update-fixs. Since #272 every app on a channel runs scripts/update_fixs.{ps1,sh}, so a broken installer breaks all of them at once.

The installers can't install from a local file, so a test before publishing had nothing to run. They now take:

• --zip <path> / -Zip: install a build zip already on disk instead of downloading one. Everything after the download runs unchanged: unpack, native runtime fetch, version marker. Behaviour without the flag is unchanged.
• --list / -List: print the releases the picker would offer, then exit.

release.yml then checks the install at two points:

When	Where	What it exercises
Before publish	build-publish (Windows), linux-bundle (focal)	installs the zip just packed with --zip. A zip that can't install never reaches the channel: the Windows step sits ahead of the publish step, and linux-publish needs linux-bundle.
After publish	new install-check-windows, install-check-linux jobs	installs the channel through the real download path, including the checksum. The Linux job also checks that --list equals the releases that carry a fixs-build-*.zip, compared against jq's reading of the API.

Both points check the result with scripts/check_fixs_install.{ps1,sh}:

• TrafficLayer is present.
• The native runtime loads: Test-LibsumoLoadable on Windows, ldd with no LD_LIBRARY_PATH on Linux.
• Line 1 of FIXS_VERSION.txt carries the tag, with no BOM.

The after-publish check can't prevent a bad release, only make it visible. Prevention comes from the before-publish check.

Checksum rule. #396 (#204) started publishing .sha256 files. A .sha256 that the release lists is now required: if it can't be fetched, or doesn't match, the install fails. A release that lists none was published before #204 and still installs, with a warning, as before. This closes the gap for new releases without breaking installs of the pinned old ones.

Build gate. scripts/update_fixs.* was on the build-skip list, so a PR touching only the installer ran nothing. It's off the list now, because the build is what installs the fresh zip with it.

#248: nothing built the env from environment.yml

New python-env.yml: create the env from the spec alone, import carla_env_setup.CHECKED_MODULES, byte-compile Carla/ and CommonLib/, add carla==0.9.15 as ensure_carla() does, and run tests/Python/unit. It has no path filter on purpose: release.yml skips Python-only PRs, and those are the PRs this workflow is for.

• Windows runner, not Linux as [MAINTENANCE] CI never tests that the shipped environment.yml produces a working env #248 sketched. The spec pins vs2015_runtime/vc14_runtime, which conda-forge publishes for win-64 only (conda search --platform linux-64: no match), so the spec does not resolve on Linux.
• traci/sumolib are left out of the import check, and the step says so in its output. Where they come from is [BUG] Shipped environment.yml does not produce a working co-sim env (carla wheel path + missing/unpinned SUMO clients) #221.
• Proposed as advisory at first. Make it a required check after a few green runs, so an environment.yml problem nobody knows about yet doesn't block every PR.

Related Issues

Closes #273
Closes #248

Environment

• Python version: 3.10 (env created from environment.yml)
• SUMO version: 1.22.0 (libsumo from fixs-native-deps)

Checklist

• Code compiles/runs as expected: locally, see below
• Tests pass locally: 451 passed, 8 skipped, in an env built from environment.yml
• Documentation is updated (if applicable): usage is in the script headers
• Issue linked above

Additional Notes

Verified locally (Windows 11, against the live releases)

update_fixs.ps1, and update_fixs.sh under Git Bash, where it selects the Windows assets:

Case	Result
--zip with the published v0.9.1-alpha zip	installs; check passes
v0.9.1-alpha (lists .sha256)	checksum OK on both zip and libsumo; check passes
v0.9.0-alpha (lists none)	warns, installs
--list	the 5 releases with a build zip; excludes fixs-native-deps, Binaries-*, v0.6.0, v0.7.0; equals the jq reading

Each negative control fails for its own reason:

Negative control	Result
zip with TrafficLayer.exe removed	check fails: TrafficLayer.exe is missing
wrong tag	check fails on line 1
sidecar forced "listed" on v0.9.0-alpha	install fails: lists … .sha256 but it could not be fetched
tampered expected hash	install fails: checksum mismatch, and no FIXS_VERSION.txt is written
--list compared against all releases	differs, as it should
env import check under an interpreter lacking pandas/shapely/networkx	fails, naming each
a .py with a syntax error under Carla/	byte-compile fails

Not verified yet

• check_fixs_install.sh and the Linux installs. They need a Linux runner (ldd), so this PR's own linux-bundle run is their first run.
• The after-publish jobs. They only run when a branch with a channel publishes. A workflow_dispatch with publish=true on this branch would not exercise them either: the branch has no channel, so nothing publishes and they skip. Their first run is the first push to dev_v0.9.1 after merge. The before-publish checks, which are the ones that block a bad release, did run and pass on this PR.
• setup-miniconda with Miniforge against a spec that lists defaults. Locally I used Miniconda. This PR's python-env run is the test.

Out of scope, found on the way

1. environment.yml's trailing comment says ensure_carla() also installs eclipse-sumo/traci/sumolib. It doesn't: only the no-conda fallback does (carla_env_setup.py:776). This belongs to [BUG] Shipped environment.yml does not produce a working co-sim env (carla wheel path + missing/unpinned SUMO clients) #221.
2. The spec does not resolve on Linux (the win-64-only runtimes above), so a Linux consumer can't conda env create -f environment.yml.
3. Carla/run_cosim.py:1605 and :1641 have invalid escape sequences ("virenv\mainvircarla.py", f"apps\{app_id}\\"). Today they are a DeprecationWarning; from Python 3.12 they are a SyntaxWarning.