[derek73]

What

Strict-gap batch B, decided on 2026-11-03: fil.toml takes de, del, san, la and los, the Spanish-derived particles of Filipino surnames (De La Cruz, Del Rosario, San Juan, De Los Santos). It held dela alone of these, so strict fil read the rest as name words. Each entry copies es.toml's kind and properties; de and los state may_be_given = false, since no file holds a given reading of either.

Measured

• Strict losses (tools/differential/config_diff.py gaps on master f48c88a). Strict fil broke 473 records on de, 94 on del, 70 on san, 46 on la and 19 on los (Helbert De La Cruz, Joshua Del Prado, Raul San Diego). With batch A's jr, the language's 1,066 broken records fall to 28 (simulated with all five batches).
• Default and plain choices. The compiled default is unchanged. Over a grid of 21 Filipino names under the three orders (each comma shape; cased, upper and lower spellings; a title and a generation beside them), the default and plain fil move no cell.
• Strict fil. Strict fil, and strict fil + en, move 50 cells each, toward the default's reading, with one class apart:
  • del, la and san hold a given reading in es.toml alone. Strict fil drops it and keeps fil's particle reading, so there they read as never-given particles.
  • A bare Del Rosario reads family 'Del Rosario', where the default reads given 'Del' and reports particle-or-given.
  • _reading_index.strict_ungated's reviewed table gains fil: del, la, san, as it holds it's for the same three.
• Committed corpora. The snapshot matches. Reading.languages of de and la now names fil among the files, so the release log, usage.rst, _types.py and two tests list it.

decisions/vocabulary-model gains the batch's entry.

Review round

An independent review found no defect. It reproduced every strict-loss figure, and found plain fil reading 0 of 61,947 local names differently. Fixed after it:

• The decisions entry states the class's comma form (Santos, Del → family 'Santos Del' under strict fil, 26 names of a 1-in-4 sample, mostly FEC) and its measured Filipino cost: 2 records (La Reina Eunice Santos, La Trang), none for del or san.
• las (3 Philippine records) is recorded as left out.
• The entry gains a recompute recipe with the grid's names.
• STRICT_UNGATED is back in alphabetical order.
• Its comment and the 3j entry date their counts (16 words, 27 pairs over 13 tags on 2026-10-31), which the strict-gap batches move. C and F carry the identical edit.

This batch's entry points to batch A's entry (#1030): merge #1030 first.

Checks

• uv run pytest -q: 35,787 passed. uv run mypy, uv run ruff check.
• sphinx-build -b doctest and -W -b html; snapshot.py.
• Not run: -n0 -m serial and Python 3.15.

The batches (A #1030, B, C, D, F) are separate PRs off the same master. Each regenerates nameparser/config/_index.py, so whichever merges later rebuilds it. This batch's decisions entry points to batch A's umbrella entry, which lands with #1030.

🤖 Generated with Claude Code