refactor: narrow the mode-specific inputs in compute_lmm_chunk_numpy

[michael-denyer]

Continues the pyrefly burn-down. 113 → 107 entries. Clears the last 6 in
lmm/compute_numpy.py.

1. Optional inputs never narrowed — 4 errors

logl_H0 and Hi_eval_null are optional because only some LMM modes take them. They were
guarded up front:

if lmm_mode in (2, 4) and logl_H0 is None:
    raise ValueError(...)

but the branches that read them test lmm_mode == 2, and nothing connects in (2, 4) to
== 2. So all four reads still saw float | None / ndarray | None.

The guard now sits at the point of use, through _require_mode_input, tying each check to
the branch that actually needs the value. Two behavioural details preserved deliberately:

• Mode 4 binds both inputs at the top of its branch, before the Score step — so an
  omitted argument still fails before any compute runs, not partway through.
• Mode 4 checks logl_H0 first. With both absent, that is the one reported.
  testcompute_lmm_chunk_numpy_missing_args_raise pins this ordering. My first attempt
  checked Hi_eval_null first and the test caught it; I restored the production order
  rather than rewriting the test.

The helper is named _require_mode_input, not _require, because the module already has a
_require(symbol, what, abi) for binding C symbols.

2. result.update(WaldResult) — 2 errors

A TypedDict is not a Mapping[str, ndarray | None] — its value types are per-key, so the
update overloads reject it. dict(wald) does not help; it degrades to dict[str, object]
(checked, not assumed). The two sites now go through _store_wald, which copies the five
keys by name. Side benefit: it makes explicit which keys the Wald path fills, which the
blanket update hid.

The _compute_lrt_numpy / _compute_score_numpy updates are untouched — they return plain
dicts and always type-checked.

Verification

This is the LMM compute dispatch, so it got a full numerical sweep rather than a spot check.

Compared (before vs after)	Result
.assoc.txt for modes 1/2/3/4 × batch and streaming backends — 8 runs	byte-identical (sha256)
pve_estimate, pve_se, n_snps_tested for all 8	identical
Guard sweep: all 16 combinations of mode × (neither / logl_H0 only / Hi_eval_null only / both)	identical outcome and identical message text in every cell, including which key set each mode populates
pytest tests/	2241 passed, 10 skipped, 3 xfailed
ruff check / ruff format --check	clean

The guard matrix, for the record:

m2_neither    ValueError: logl_H0 is required for LRT (mode 2) and All (mode 4)
m3_logl_only  ValueError: Hi_eval_null is required for Score (mode 3) and All (mode 4)
m4_neither    ValueError: logl_H0 is required for LRT (mode 2) and All (mode 4)
m4_hi_only    ValueError: logl_H0 is required for LRT (mode 2) and All (mode 4)
m4_logl_only  ValueError: Hi_eval_null is required for Score (mode 3) and All (mode 4)
m4_both       ok:betas,lambdas,lambdas_mle,logls,p_lrts,p_scores,pwalds,ses

Baseline

Regenerated on Python 3.11.13 / numpy 2.4.6 per #121, verified on both:

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

Running total

174 (#121) → 152 (#122) → 135 (#123) → 113 (#124) → 107.

src/ is now down to 18 entries across 5 files. Everything else is tests and scripts:
tests/test_incremental_writer.py (23, still undiagnosed and the single largest cluster),
tests/test_jlinalg_dgemm.py (8), lmm/io.py (7 — an unguarded file handle plus an
unbound-name), tests/test_kinship_io.py (7).