docs: state what the doc-anchor gate does not prove, and name both absences

[michael-denyer]

Two small things, neither changing behaviour. Closing out the two remaining items from the thermo-nuclear review, one by fixing it and one by recording it rather than fixing it.

The gate's guarantee was overstated by omission

check-doc-anchors.py has to work out which symbol an anchor means. For CODEMAP's tables that is a guess: the rows label the link file.py:123 and put the symbol in a separate column, so _wanted_symbol takes the first plausible backticked name in the row. A row naming two symbols can be checked against the wrong one, which passes when that other symbol happens to sit on the anchored line.

That was documented nowhere a reader would hit it. _wanted_symbol's docstring explained the fallback mechanism without ever saying the choice is a guess, and DEVELOPMENT.md listed the hook among the prek gates and said nothing about its guarantee.

That mattered more after last week's work. The gate now sits under 94 anchors and caught real rot twice during the 7.2.0 doc pass, so its reputation had started running ahead of what it can prove. Both the module docstring and DEVELOPMENT.md now say it plainly: a green run means no anchor is provably wrong, not that every anchor is provably right.

The fix is not in this PR, deliberately. Labelling each link with its own symbol makes the label authoritative and the row irrelevant, but that is roughly 120 links across the docs. Not worth it today. Recording the limitation is, because the alternative is quietly relying on it.

Two absences, now both named

load_gemma_assoc read optional columns through _opt_float and then, on the next two lines, beta and se through an inline float(row.get(name, "nan")). Two idioms side by side for what looked like one idea.

It is not one idea. An absent optional column is None, because the test does not report it. An absent beta is NaN, because AssocResult requires it and GEMMA's LRT formats omit it. So this is not one helper with the duplicate deleted; it is _opt_float and _float_or_nan, with the difference stated rather than inferred from which idiom the author happened to reach for.

The gate caught me

_float_or_nan cost 12 lines, which shifted three CODEMAP anchors into compare.py. check-doc-anchors failed and named all three. Corrected here. That is the third time in a week it has caught a shift a human would not have, which is also the argument for being honest about where it cannot.

Verification

Same pin as #152: compare_assoc_results over all seven committed .assoc.txt fixtures, four conditions each, digesting the rendered AssocComparisonResult with skip messages included. All 28 digests identical to master, so the helper swap changed nothing.

• pytest tests/: 2275 passed, 7 skipped, 3 xfailed
• prek run --all-files: all hooks pass, check-doc-anchors at 94 OK
• pyrefly check: 0 errors, 43 suppressed, unchanged

Not done, and I would leave it

The third item from that review was extracting the .fam parsing subtree out of pipeline.py. I would drop it rather than carry it. The only reason it was on the list is that the file was over 1000 lines, and #151 took it to 897. parse_phenotypes is monkeypatched as a runner method in four places in test_pipeline_helpers.py, so the extraction costs real test churn to buy cohesion in a file that already reads fine.