fix: CI never ran tests for 3 of 4 packages, or even triggered on them

[musicalplatypus]

Problem

Multiple independent peer reviews of PRs opened this session flagged the same gap repeatedly: none of the new regression tests added to tinyml-tinyverse, tinyml-modelzoo, or tinyml-modeloptimization/torchmodelopt actually run in CI. Two compounding problems:

1. .github/workflows/test-modelmaker.yml's push/pull_request path filters only watched tinyml-modelmaker/** — a PR touching only tinyml-tinyverse, tinyml-modelzoo, or tinyml-modeloptimization (which is most of this session's PRs) never triggered the workflow at all.
2. Even when the workflow does run, it only ever executes tinyml-modelmaker/tests/ (four specific files via named Tier 1/2/3 steps). The other three packages have no test-running step whatsoever.

Fix

• Restored the tinyml-tinyverse/**, tinyml-modelzoo/**, and tinyml-modeloptimization/** path filters alongside tinyml-modelmaker/**, so the workflow actually triggers for PRs touching those packages.
• Added a test step for each of the three previously-uncovered packages, guarded by a find-based existence check (not a plain test -d tests) since none of these tests/ directories exist yet on main — an empty/missing directory makes pytest exit non-zero ("no tests collected"), which would otherwise fail CI on this package until the first PR adding tests to it merges. The guard lets this land safely now and starts actually running tests the moment any PR adds them.

Deliberately out of scope

I considered also adding a broader "run everything under tinyml-modelmaker/tests/" sweep, since several existing test files there (test_constants.py, test_dataset_utils.py, test_nas_support.py, etc.) aren't wired into any of the three named Tiers and are therefore also never run. I dropped that: it surfaces 17 pre-existing failures unrelated to anything in this session (confirmed even the current, unmodified Tier 1 step alone already fails 2/447 tests locally), which would immediately turn this workflow red on merge. That's a real, separate gap, but out of scope for a fix that should land safely without breaking CI.

Testing

Verified locally: the find-based guard correctly skips tinyml-tinyverse, tinyml-modelzoo, and torchmodelopt on this branch (no test content exists yet on main) and would correctly run pytest once a PR adds real test files there. YAML syntax validated.

🤖 Generated with Claude Code

[musicalplatypus]

An independent peer review of this PR caught an in-scope gap. Pushed a follow-up commit (9775450):

The pull_request path filters omitted .github/workflows/test-modelmaker.yml itself (the push filters have always had it -- an inherited asymmetry). Consequence: a PR that only changes this workflow, including this very PR, never triggers the CI it modifies, so workflow changes merge unvalidated. Given this PR's stated purpose is "CI doesn't trigger when it should," that belongs in scope. One added path glob, mirroring the push filter.

The same review also verified the load-bearing claims: for pull_request events GitHub evaluates both path filters and the workflow file from the PR's merge ref (base+head), so once this merges, the other open PRs pick up the new triggers/steps on their next synchronize event with no rebase -- though merging this does not retroactively spawn runs, so each open PR needs one push/synchronize to get its first run under the new workflow. Dependency coverage for the not-yet-existing test suites was also verified (tinyml-tinyverse's pyproject declares onnxruntime/torcheval/etc. and is installed without --no-deps, and every new test file across the open PR branches imports only stdlib + already-installed packages).