[derek73]

A follow-up to #902 (phase 2m of #752). Over three review rounds, the code that names Ambiguity.alternative's field grew into five helpers that each re-derived part of one question: where would this word land as a name word? This PR merges what they shared and gives PendingReading's BESIDE-only data its own type. No reading or alternative moves.

What changes

• _assign._slot_field(state, seg_idx, pos, count, order) answers which field a name slot takes, by structure:

  • the family before a family comma;
  • the given part's given or middle name after one;
  • otherwise, the order the main segment's read set.

  as_a_name, _split_name_alt and _resolve_deferred each had their own copy of this; they now call it.

• _assign._name_pieces(state, tokens, seg_idx) lists a segment's name pieces, for _split_name_alt, _resolve_deferred and _post_rules._unattached.

• Beside(index, unfolded) in _pipeline/_state.py replaces Deferred.BESIDE and the two optional PendingReading fields beside and unfolded, which meant something only under it. PendingReading.role is now Role | Deferred | Beside, and _assemble._alternative reads the variant.

• _name_alt's parameter family_fixed is renamed licensed. It is P5's (bound given names) join licence, which a given-name title grants as well as a family comma.

Left alone, as scoped: the snapshot and compare tooling, and as_a_middle and the P6 report site in the given-part walk. Those two look up the piece in front by raw index, which is a different question from _name_pieces. They already name the middle slot _slot_field would give them there.

Measurements (2026-10-26)

• No alternative moves. I recorded every report's kind and alternative (reading kind, role, source) over 542,555 parses, before the change and after. The grid:

  • the case table under all three orders;
  • _REVIEW_SHAPES;
  • the snapshot corpora under six policies (the three orders, middle_as_family with and without FAMILY_FIRST, and two_family_names);
  • a generated grid of every one- to three-word run from a 21-word pool (particles, bound given names, numerals, credentials, titles, a marker, the one-case connective) in nine comma shapes.

  Of 234,112 reports, 232,739 carry an alternative, and the two records are identical. Negative control: with the given part's first slot mutated in _slot_field, the same record moves 198.

• Snapshot and local corpora. The committed snapshot matches. snapshot.py --local tse-donors and --local ctgov-officials with --base master match with alternatives compared, since both sides' headers record them.

• Frames. The reference row keeps 305 and Smith, John 158, and the serial benchmark passes. A name that reports pays one frame per _slot_field call, and a connective's report one more for _name_pieces: JOHN SMITH MA 223 → 224, MARIA DA SILVA E SOUZA 294 → 296. The cost is per report, not per token, and no guard moves.

• Checks. uv run pytest passes (31,529), as do the serial suite, Python 3.15, mypy, ruff, the sphinx doctest build and the README doctest.

• Review. The code review found no defects. It flagged two lines the refactor left unwrapped, fixed in 80f64af. A second review checked that commit and found it is wrapping only.

Refs #752.

🤖 Generated with Claude Code