pyright: 15 errors to 0, by narrowing where the code already guarantees it

[soodoku]

Every one of these was code that is correct today and could not be shown to be. No behaviour changes.

pred_census_ln.py — 4

The validation raises when stats is None and either prior or coverage was given, and every use of stats sits inside one of those two conditions — a chain no checker follows through a compound and:

if (prior is not None or coverage is not None) and stats is None:
    raise ValueError(...)
...
if coverage is not None:
    qhat = stats["conformal_quantiles"][...]   # pyright: stats may be None

Asserted at the three use points, which also states the invariant for the next reader. The coverage validation is now a nested if rather than a compound one, which reads better anyway.

pred_wiki_origin.py — 2

The base declares VOCABFN/RACEFN as str | None because non-LSTM models have neither; this model sets both.

Narrowing them in the subclass does not work — a mutable ClassVar is invariant, so str is not assignable to str | None. I tried it and it traded two errors for two different ones, so it is narrowed at the call site instead.

dict_models.py — 9

pandas-stubs types df[key] as Series | DataFrame, because a column label could be duplicated. A scalar key yields a Series and a list key a DataFrame — which the stubs cannot express. Five Series call sites and one DataFrame assignment are cast, each with the reason.

census_marginal returns a cached value the branch above always sets, so that is asserted rather than cast.

Verification

pyright 15 → 0, ruff check and ruff format --check clean, 211 tests pass.

Combined with the earlier scoping PR, this repo goes from 130 pyright errors and 12,249 codespell findings to zero of each.