Deferred CI follow-ups dropped when #32 auto-closed

[d-morrison]

Recovering work dropped when #32 auto-closed.

#32 carried a "Deferred" section listing four follow-ups. Its PR (#33) said
Closes #32, so merging closed the parent and took the deferred items with
it. GitHub cannot partially close an issue, and a closed issue reads as
evidence the work was handled --- so these were not deferred, they were
silently gone.

Re-filed here, unchanged in substance.

1. Migrate test-coverage.yaml to Morrison-Lab/gha's test-coverage.yml@v2

Blocked on the workflow's CmdStan install and cache steps, which gha does
not model.

Those steps look vestigial on main: no .stan file is tracked there and
cmdstanr is not in DESCRIPTION. But #13 is an open Stan PR and #31 carries
.stan files under inst/extdata/, so removing them could break coverage the
moment either lands. Settle that first.

2. Add check-non-standard-chars.yml

Would fail today on em-dashes in .claude/skills/reprexes/SKILL.md (24 lines)
and .github/copilot-instructions.md (16 lines). The workflow has no
paths-ignore input, so those two files need cleaning first.

Note that both are AI-instruction files where the em-dashes are deliberate
prose style, so this is a real decision rather than a mechanical sweep.

3. Flip the three whole-tree linters from warn-only to blocking

lint-workflows.yml, lint-yaml.yml, and lint-markdown.yml were added at
fail: false in #33, because the repo had pre-existing drift.

One trap when doing this: lint-markdown's fail input gates markdownlint
alone. fail-on-item-splices and fail-on-table-splits already default to
true and already block, while fail-on-long-code-blocks defaults to
false. So "flip it to blocking" is narrower than it sounds --- check each
companion's own default rather than assuming one input governs the workflow.

4. Consider check-ai-tells.yml

Needs threshold tuning against this repository's research prose before it
would be useful rather than noisy.