Decide object, module, type, list and float

[tamnd]

The rows behind tamnd/python-docs-vi-translator#58, now that a row can say two things.

What was measured

Over the 548 catalogs of the pinned mirror, counting only prose the glossary matcher finds the term in, and masking markup on the Vietnamese side so :class: list`` does not read as the word "list":

term	Vietnamese	English	decision
object	đối tượng 63	object 46	đối tượng
list	danh sách 34	list 15	danh sách, stands alone
type	kiểu 30	type 16	kiểu, stands alone
float	số thực 4	float 4	số thực, stands alone
module	mô-đun 17	module 64	module, kept

type, list and float stand alone because each of them names a thing in a table somewhere: type is the argparse keyword in a table of argparse keywords, list is a cell in the json conversion table next to array, float is a cell in the struct format table next to _Float16. object does not, because 35 of its 36 standalone entries are the index category and read "bytearray (object)" to somebody using the index.

module had no row at all and no majority either, 64 to 17, and the split does not resolve by area: the tutorial runs the other way from the library reference. So it is decided on the reason rather than the count. mô-đun is module spelled with a hyphen in it and carries nothing a reader does not already get from module, where đối tượng, kiểu, danh sách and số thực are Vietnamese words that mean something. bytecode, docstring, iterable, lambda, traceback and REPL are already kept for exactly that reason.

import, slice and callable are the same shape as float and are deliberately left as they are. Their Vietnamese moves with the part of speech, nhập for the verb and import for the noun in "hệ thống import", and a row naming one rendering would fail G02 on the other. They stay keep_en until somebody words them.

What it moves

check	before	after
L02 untranslated (hard)	86	45
G03 kept term lost	94	103
G02 rendering missing	191	339
G04 term left in English	143	313

41 hard findings for 327 soft ones, and the soft ones are the point. Of G03's 103, module is 80, and 65 of those 80 are machine output from the 2026-08-17 tier 1 run. The model was inventing mô-đun because the glossary said nothing about module, and now it will not. The same holds across the new G02 and G04 findings: 198 of the 362 are machine, and the next run clears them.

The other 164 are human translations that disagree with each other, which is what a glossary is for and what a report is for. reports/quality.md now reads 0.19 misses per machine entry against 0.05, and the reason it read 0.05 is that the glossary had nothing to say about the words the model was getting wrong.

L02's remaining 45: object 36, Lists 3, three copyright notices, a Monty Python transcript and two one-offs. Lists is not folded into list on purpose; it is a section heading three times, each followed by prose beginning "Lists are mutable sequences", and a heading is translated.

Audit

548 catalogs, 41 checks, 29 passing, 2,545 findings. Hard failing: P01 26, P03 27, L01 4, L02 45.

Glossary v2 to v3, 642 terms, 51 kept in English. Every catalog rewritten for X-Generator: pydocvi 0.1.11.