ci: the path-aware skip covers only the Companion

[L4XB]

Problem

.github/workflows/ci.yml has a changes job that detects which areas a pull
request touches, but it emits exactly one output:

    outputs:
      companion: ${{ steps.areas.outputs.companion }}

Only the macOS Companion jobs use it. Everything else runs in full for every
pull request — the API suite with its ~2,900 tests, the web build, the container
builds — including on a change that touches one markdown file.

The cost is not abstract. Both branches require checks to be up to date before
merging, so a pull request that goes stale has to re-run the whole matrix. A
documentation-only change currently pays the same ~15 minutes as a change to the
API, and while it runs, every other pull request behind it is invalidated by the
merge.

What to do

Extend changes with outputs for the other areas and gate the expensive jobs on
them, following exactly the pattern the Companion jobs already use.

The trap to avoid

The required status checks are configured by name. A job that is skipped
outright never reports its check, and the merge blocks forever waiting for it.
That is why companion-macos does not skip — it falls back to
ubuntu-24.04 and reports success without doing the expensive work:

runs-on: ${{ needs.changes.outputs.companion == 'true' && matrix.os || 'ubuntu-24.04' }}

Any new gating has to keep reporting the check. Copy that shape rather than
adding if: at the job level.

Careful about what counts as "touched"

services/api changes affect the web contract test. A lockfile change affects
everything. Err toward running: a skipped job that should have run is a bug that
reaches main, while a job that ran unnecessarily only costs minutes.

Acceptance criteria

• Every required check still reports on every pull request
• A documentation-only change measurably shortens the run — state the before
  and after in the pull request
• The trigger set for each area is listed in a comment, as the Companion one is
• A change to a shared file (a lockfile, ci.yml itself) runs everything