fix(flint-py): point run_full_eval at the moved fixture corpus

[zl190]

Summary

• tools/run_full_eval.py exits with ERROR: manifest.json not found and evaluates nothing: FIXTURES still points at packages/flint-py/tests/fixtures/, but the corpus moved to shared/test-data/ in 1d1ee39.
• Repointing FIXTURES on its own would have dragged the report along with it, since line 345 derived that path as FIXTURES.parent — hence the separate REPORT_DIR.
• Repointing the corpus and giving the report its own constant brings the tool back; it writes results.json and FULL_GALLERY_REPORT.md to the two locations where the committed copies already live.

Context

Found while investigating #80. The eval tool is the instrument that measures flint-py against the JS reference, so having it run again is the prerequisite for acting on anything in that issue.

Changes

packages/flint-py/tools/run_full_eval.py (+8 / −4):

• FIXTURES → shared/test-data/, with a comment naming the commit that moved it.
• New REPORT_DIR = ROOT / "tests" so FULL_GALLERY_REPORT.md keeps landing next to its committed copy rather than following the corpus.
• Module docstring updated to name both output locations.

Review focus

1. Are those the two output locations you want? results.json goes to shared/test-data/results.json (which is tracked there already) and the report to packages/flint-py/tests/FULL_GALLERY_REPORT.md (likewise). I inferred the intent from where the committed copies sit — say the word if either belongs elsewhere.
2. Should the regenerated outputs be part of this PR? Running the tool rewrites both files. I left them out so the diff stays at the fix. See the note below on what they now say.

Test plan

• uv run python tools/run_full_eval.py from packages/flint-py — completes and prints: PASS: 476, MISMATCH: 180, PY_ERROR: 0, JS_ERROR: 1, FIXTURE_MISSING: 0.
• Confirmed the two outputs land on the committed copies (git status shows exactly packages/flint-py/tests/FULL_GALLERY_REPORT.md and shared/test-data/results.json as modified, nothing untracked).
• Cross-checked the tool's MISMATCH: 180 against uv run pytest -q, which reports 180 failures. Two independent instruments, same number.
• Reverted both regenerated artifacts, leaving the diff at the one file.

Notes for reviewers

• What the refreshed report says. Regenerating it turns the summary from 658 / 658 (100.0%) into 476 / 656 (72.6%). That drop is what Golden fixtures in shared/test-data no longer match the JS reference (705/705) #80 is about; it is not caused by this change, which touches only where the tool reads and writes.
• Small pre-existing inconsistency, not fixed here. The report generator prints "There are N specs that differ … i.e. the Python port matches the JS reference on every ported chart type for every gallery test case" as one unconditional sentence, so with N > 0 it contradicts itself. Left alone to keep this diff to the paths; happy to send a follow-up.

[Chenglong-MS]

this is great, the python library is quite outdated, I think it's good that we should use the new test cases to keep them in sync.