improvement: re-check the subproject PLT outside mix check's parallel section

[jimsynz]

Fixes #245 together with beam-bots/bb_example_wx200#97 and
beam-bots/bb_example_so101#91.

Root cause

Not the cache key, and not a stale PLT. mix gettext.extract --check-up-to-date
re-expands the gettext macros by force-recompiling the project
(mix compile --force-elixir + mix compile.elixir --force), and a forced
recompile empties _build/dev/lib/<app>/ebin and .../consolidated before
refilling them. Polled every 50 ms in bb_example_wx200, that window is about
two seconds:

40.86 ebin=21 consolidated=27
41.59 ebin=21 consolidated=0     <- consolidated/ emptied
41.70 ebin=1  consolidated=0     <- ebin emptied, .app file only
43.47 ebin=21 consolidated=0
43.76 ebin=21 consolidated=27

ex_check runs the gettext tool in parallel with dialyzer. Elixir's build
lock serialises the two compiles — that is what the "Waiting for lock on the
build directory" lines in the failing logs are — but not dialyzer's beam reads,
which happen after it releases the lock: first the PLT check, then the
Path.wildcard in Dialyxir.Project.dialyzer_files/0 that builds its file list.
A wipe landing across both gives exactly the reported failure: the PLT check
can't read a consolidated protocol beam, and the file list comes back empty.

Only bb_example_wx200 and bb_example_so101 carry both gettext and
dialyxir, which is why only those two jobs fail — bb_example_so101 failed
identically on the v0.30.1 tag push (run 32619948357), so it was never really
PR-only either. The issue's consolidate_protocols lead is a red herring: those
two are Phoenix apps and don't set it, so they do consolidate in :dev. Setting
it would remove one of the two error lines and leave the empty file list.

What this changes

The Force a PLT re-check step deleted _build/*/*.plt.hash and left the
re-check itself to mix check, where it runs alongside every other tool and
spends twenty-odd seconds reading every beam in the build directory. That is the
amplifier: with a valid hash, dialyzer's exposure is the ~0.15 s between
releasing the build lock and globing the ebin; with the hash deleted it is the
whole PLT check. This PR primes the PLT in that step instead, so the re-check
still happens — same correctness, same total work, moved out of the parallel
section.

That narrows the window; it doesn't close it. The two .check.exs PRs above
close it, by making the gettext tool depend on dialyzer so the forced
recompile runs when nothing else is reading _build. All three are needed.

The first commit is the real source touch used to make the downstream matrix
recompile bb (a one-character typo fix in the color entity's docs, plus the
regenerated cheat sheet). It's a genuine fix, so it stays rather than being
reverted.

Verification

Reproduced locally, byte for byte. In bb_example_wx200 with
BB_VERSION=local and _build/*/*.plt.hash deleted, overlapping
mix gettext.extract --check-up-to-date with mix dialyzer:

:dialyzer.run error: File not found:
  .../_build/dev/lib/bb_example_wx200/consolidated/Elixir.Collectable.beam
  files: [],
:dialyzer.run error: Analysis failed with error:
No .beam files to analyze (no --src specified?)
Halting VM with exit status 1

Same file, same empty files: [] as the CI logs. With the .check.exs fix,
mix check --no-retry moves the forced recompile to the last three seconds of
the run, after dialyzer has finished.

Not reproduced naturally in CI, and I can show why. The bb_example_wx200
and bb_example_so101 jobs pass on this PR — but they also pass on a scratch PR
carrying only the source touch with this workflow unchanged (#247, closed), so
that green is not evidence of a fix. Re-running the 23 Aug job that failed
(run 32674375855, job 97280389283) unchanged, today, also passes. The timings
against the same _build cache key have moved a long way since:

	23 Aug (failed)	today (passes)
ex_unit	0:13	0:49
dialyzer	0:21	1:51

gettext is gated on ex_unit, so it now starts around 50 s in, well clear of
dialyzer's PLT check. That is consistent with a timing race rather than anything
about the branch, and it means CI cannot currently demonstrate either the break
or the fix.

This PR's step does behave as intended (job 98350500869): the PLT check runs
in its own step for 81 s, and mix dialyzer inside mix check then drops from
0:21–0:36 to 0:16 with no PLT-check phase at all.

[jimsynz]

reopening to kick off the workflows