fix(dataset-tests): route DatasetTestRunner through the shared output emitter

[sakanni]

DatasetTestRunner carried its own copy of the shared output emitter, and the copy had drifted. It omitted title= and the location prefix on the message, which is the workaround the shared emitter documents for GitHub stripping file= out of the rendered log line. So a failing dataset fixture produced an annotation that anchored correctly on the diff while the job log line named no file at all, and the log is what people actually read. It also mapped every non-failure level to warning, so a notice was reported as a warning, and it flattened newlines to spaces rather than %0A, so a nested result hierarchy arrived as one long line.

Routing through OutputEmitter.Write fixes all three. Write already takes an optional FileAccounting, so the same call site also gives this runner the coverage denominator the other runners report. There is no relevance filter in this runner, so NotRelevant stays zero by construction.

The second commit stands alone deliberately. The two [SKIP] diagnostics were written unconditionally to stdout, which corrupts --output json and --output sarif, since those put a single payload on stdout and nothing may precede it: one missing fixture made the output unparseable from its first character. This is pre-existing, and the shared emitter already cites this diagnostic as the reason it guards its own coverage line. Routing around it stopped being sufficient once the counts became part of the json payload. They now go to stderr, so the diagnostic still reaches the job log everywhere it was visible before. This affects --output json and --output sarif consumers only; production passes --output github and is unaffected either way.

Verified on a Windows runner against a repository carrying dataset fixtures, with one fixture's expected output deliberately made wrong so the runner had a finding to emit, and with annotations read back from the check-runs API rather than from the log. The annotation gains a title carrying its location, the rendered log line gains the path prefix, and a coverage line appears; path, start line and annotation level are unchanged. The verdict is unchanged, checked on the mutated fixture and again across every fixture in that repository, with identical failure sets before and after.

The project is added to Platform.slnx, which is what makes the tests possible: it was previously compiled only by prepare-runner at check time, so no solution build covered it and nothing could invoke it. The emitted annotation shape is already covered by the existing emitter tests, so the new tests do not restate it; they cover that this runner routes through the shared emitter at all, and that accounting is wired at each exit the loop can take. The suite is green at 64 tests against 59 before.

Not verified: the notice-to-warning mapping is a source reading. No notice-level finding arose in any run, so I have not observed that path.