refactor: clear the 22 pyrefly baseline entries in validation/compare.py

[michael-denyer]

Continues the pyrefly burn-down from #118 / #121. 174 → 152 entries.

Two root causes, both in src/jamma/validation/compare.py.

1. The dict-splat — 15 errors at one call site

load_gemma_assoc built a row dict through _parse_assoc_field, which returns
str | int | float, then splatted it: AssocResult(**row). That union reaches every
parameter, so all 15 fields error at the single line 299.

Replaced with explicit construction: raw cells stay dict[str, str], and each field gets
its own int() / float(). Six float | None columns go through a new _opt_float(row, name)
helper that returns None when the layout omits the column. Header names already equal
AssocResult field names, so the mapping is direct — verified, not assumed: every column
across all seven accepted layouts is an AssocResult field, and every field appears in at
least one layout.

_parse_assoc_field, _ASSOC_STR_FIELDS and _ASSOC_INT_FIELDS had no other callers and
are deleted.

2. The duplicated dispatch — 7 errors

compare_assoc_results ran the same four-way test-type dispatch twice: once to compute the
per-column ComparisonResults, once to fold them into passed. The second chain read
p_score_result.passed, p_lrt_result.passed and l_mle_result.passed with no narrowing,
because the assignments happened in the first chain — 6 missing-attribute errors. The 7th
was passed=all_passed, which picked up numpy.bool from the LRT branch's np.all(...).

Merged into one dispatch so each branch computes its columns and its verdict together, and
cast the two np.all results to bool.

That also removes a latent AttributeError. The two chains ordered the branches differently
— first all / score / lrt / wald, second all / lrt / score / wald. is_score_test and
is_lrt_test are not structurally exclusive (both require not is_all_tests and all
p_wald is None), so a result set satisfying both would compute the Score columns and then
read p_lrt_result.passed on None. Unreachable from load_gemma_assoc — no header layout
carries p_score and p_lrt without p_wald — but reachable through the public API, which
takes a list[AssocResult] directly. The merge makes it structurally impossible.

Verification

Equivalence proved against the fixtures rather than argued:

Check	Result
Old vs new parser over all 19 .assoc.txt fixtures	identical — 90,644 rows, exact repr match (catches 1 vs 1.0 type drift, not just value drift)
Old vs new compare_assoc_results	identical — 95 cases: identity, three perturbation scales, and a length mismatch, per fixture
pytest tests/	2241 passed, 10 skipped, 3 xfailed
ruff check / ruff format --check	clean

Fixture coverage spans all four LMM modes (wald, lrt, score, all) and all seven header
layouts, including both logl_H1-present and logl_H1-absent variants.

Baseline

Regenerated on Python 3.11.13 / numpy 2.4.6 — the requires-python floor, per #121 —
and verified on both interpreters:

Environment	pyrefly check --baseline
Python 3.11.13, numpy 2.4.6	0 errors
Python 3.13.5, numpy 2.5.1	0 errors

Real backlog without --baseline: 152, down from 174.

Note for anyone repeating this: a fresh uv sync in a new worktree resolves 3.13, not
the floor — there is no .python-version in the repo. uv sync -p 3.11 first, or the #121
regression comes straight back.