One SWC display vocabulary, and a release workflow that needs no local build

[longvo92]

Two commits. The first finishes the rule-3 cleanup the previous PR deferred;
the second removes the reason a release currently costs a manual build.

7618e33 — one SWC display vocabulary

The previous PR reported three duplicated helpers between report.py and
qtviewer/. That count was low. What was actually there:

Fact	Copies	Where
iface_kind	3	report.py, summary_model.py, main.py (inlined .replace())
swc_item	3	report.py, summary_model.py, main.py
the category list	6	report.py ×3, summary_model.py, diffpane.py, main.py

main.py is a third renderer — the terminal summary — and the audit only
looked at the report/viewer pair, so it was missed. report.py alone spelled
the list out three times, in two label styles.

view_model.SWC_DISPLAY now holds the labels and takes the keys from
arxml_rules.SWC_CATEGORIES, the module that produces them. A category added
to the rules without a label raises at import instead of quietly going missing
from one surface. (arxml_rules.SWC_CATEGORIES already existed — adding a
second constant of the same name in view_model was the first thing this
change nearly got wrong.)

The blind spot worth knowing about: the fixtures never change a runnable,
so any surface could have dropped that category with the whole suite still
green. The new tests use a synthetic swc diff with all three categories
populated and assert each surface names all of them. Planting a hardcoded list
missing runnables back into _swc_note fails exactly one test.

This touches rendering code, so behaviour was checked rather than assumed: a
report and a terminal run were generated from origin/main in a temporary
worktree and compared against this branch. The terminal output is identical;
the report differs by one meta line, because the two runs were invoked with an
absolute and a relative fixture path.

bcadd35 — release workflow

Cutting a release meant building the .exe by hand and uploading it, which is
part of why the one-line fix sitting in [Unreleased] was not worth a version.

python packaging/release_check.py 1.3.1                                  # same check CI runs
gh workflow run release.yml --ref main -f version=1.3.1                  # rehearsal
gh workflow run release.yml --ref main -f version=1.3.1 -f publish=true  # for real

The default does not publish. A run without publish builds both
artifacts, runs each against the fixtures and stops — and uploads them either
way, because the point of a rehearsal is being able to look at what would have
shipped. The irreversible half is opt-in.

packaging/release_check.py owns every precondition and prints what to change,
since the reader is usually a log:

Not ready to release 1.3.1:
  - compare_tool/__init__.py says __version__ = '1.3.0', not '1.3.1' -- bump it and merge that first
  - CHANGELOG.md has no "## [1.3.1]" section -- move the [Unreleased] entries under it, with the date
  - CHANGELOG.md still has entries under [Unreleased]; they would ship without being listed in the release notes

One script, so a local check and the workflow cannot disagree. The workflow
adds the two facts a file cannot know: the ref is main, and the tag is free.
A published version is refused, never overwritten. Nothing in the workflow
edits the repo — the version bump arrives through a normal reviewed PR.

For the reviewer

• python -m unittest discover -s tests — 447 pass. python -m ruff check . clean.
• The build steps were run locally instead of being discovered on the first
  tag: build.ps1 -Pyz produces the 44 MB exe and the zipapp, both exit 1 on
  the fixtures, and the artifact-verification step was executed against them —
  and against two identical folders, to watch it fail with
  ::error::dist/compare-tool.exe exited 0, expected 1.
• Still unproven, by nature: gh release create only runs on a real publish,
  and actions/upload-artifact only exists on a runner. The first rehearsal
  covers both.
• No CHANGELOG entry. The refactor is invisible, and the release workflow is
  maintainer tooling.