refactor: declare real types in pipeline.py and pipeline_config.py

[michael-denyer]

Continues the pyrefly burn-down. 152 → 135 entries.

Five root causes, all the same shape: a declared type looser than what the code actually
holds, so every read of the value has to answer for a case that cannot happen.

Root cause	Errors	Fix
_run_batch(plink_data: object | None)	7	PlinkData | None
PipelineConfig.phenotype_columns: list[int] | None	4	list[int], empty-list default
JAMMA_BACKEND as unvalidated str	2	validate at the boundary
run_result = None before the phenotype loop	2	capture pve/pve_se per iteration
_PhenoLoopOutcome.runner_timing: dict[str, float]	1	RunnerTiming
PipelineResult.timing = field(default_factory=dict)	1	default_factory=PipelineTiming

plink_data: object | None — 7 errors

object has no attributes, so all seven reads (.genotypes, .n_snps, .chromosome,
.sid, .bp_position, .allele_1, .allele_2) error. The value is always a PlinkData,
which jamma.io.plink exports and pipeline.py already imports from — no new dependency,
no cycle.

phenotype_columns — 4 errors

list[int] | None, assigned in __post_init__ and never None afterwards, so all four
readers had to handle an impossible None. Declared list[int] with an empty-list default:
empty is now the "unspecified" sentinel. Side effect — phenotype_columns=[] derives from
phenotype_column rather than raising IndexError on self.phenotype_columns[0], which
is the behaviour the field default now implies. Passing None explicitly still works at
runtime (verified below).

This does not touch the phenotype_column / phenotype_columns duality itself — the
mutation-based sync in __post_init__ and the range check living in
PipelineRunner.validate_inputs are both still outstanding.

JAMMA_BACKEND — 2 errors

os.environ.get("JAMMA_BACKEND") is str | None, so requested widened to str and no
longer matched select_execution_mode's Literal. Rather than cast, it's validated where
it enters the process. select_execution_mode did already reject unknown values, so this
is not a new guard — but it now fires before the pipeline reads PLINK metadata off
disk, and with a message that names the env var.

The ("auto", "numpy", "numpy-streaming") triple had three copies. Two now share
BackendRequest / VALID_BACKENDS in pipeline_config; select_execution_mode keeps its
own, since collapsing that one means touching lmm/runner.py.

run_result = None — 2 errors

Initialised to None before the per-phenotype loop, then read as run_result.pve after
it. Replaced with pve/pve_se locals captured each iteration — which is what the code
meant, the last phenotype's estimate. On an empty column list the old code raised
AttributeError; it now yields the None the field already declares.

Verification

The type checker says nothing about behaviour, so the loop restructure was checked against
real runs. Single- and multi-phenotype (-n 1 and -n 1,6) on mouse_hs1940, both the
batch and streaming backends, before and after:

Compared	Result
.assoc.txt output, all 6 files	byte-identical (cmp)
pve_estimate, pve_se, n_snps_tested, n_covariates, timing keys	identical
pytest tests/	2241 passed, 10 skipped, 3 xfailed
ruff check / ruff format --check	clean

Config derivation checked directly:

unspecified                    -> columns=[1] scalar=1
empty list                     -> columns=[1] scalar=1
explicit None (untyped caller) -> columns=[1] scalar=1
scalar=3                       -> columns=[3] scalar=3
explicit [2,4]                 -> columns=[2,4] scalar=2

PipelineResult().timing is still {} at runtime.

Baseline

Regenerated on Python 3.11.13 / numpy 2.4.6 per #121, 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

What's left in this file

One entry, pipeline.py:765: compute_loco_kinship_streaming returns
Iterator | tuple[Iterator, SnpStatsCache] depending on its return_snp_stats flag.
Fixing it means @overload on the callee in kinship/compute.py, which also clears that
file's own 6 entries and lets lmm/loco.py drop a cast whose comment already claims the
function "is overloaded on return_snp_stats" — it isn't, yet. Separate PR.