L02 cannot tell a kept identifier from an untranslated word

[tamnd]

Narrowing the identifier rule in #47 put 6 558 human-written single-word entries into Corpus.prose() for the first time, and L02 went from 10 findings to 144 over 20 distinct msgids. Roughly two thirds of those are real and one third are the check being wrong.

Real, a reviewer left an English word in a Vietnamese catalog:

msgid	Entries
module	41
object	36
type	12
Boolean	4
Lists	3

Not real, a reviewer correctly kept a Python identifier:

msgid	Entries
sys	20
builtins	8
import	5
exec	3
NaN	1
Infinity	1

L02 asks whether the translation is the English verbatim, and for sys the answer is yes and it is the right answer. The classifier cannot help here, because that is exactly the discrimination #47 established it cannot make from the string alone: sys and Notes are the same shape.

The glossary already has the concept. G03 checks that a term marked kept stayed in English, so a kept entry for each of these is the statement that the corpus is missing, and L02 should not flag an entry whose whole msgid is a kept term. That turns 38 findings from noise into a checked claim, and it makes the disagreement visible: sys is hệ thống in 38 entries and sys in 20, so today the corpus says both.

Two parts:

• Add the kept identifiers to the glossary. sys, builtins, import, exec, NaN, Infinity are the ones L02 has found so far, and the same sweep over single-word entries will find the rest.
• Exempt an entry whose stripped msgid is a kept glossary term from L02, so a kept term is checked by G03 and by nothing else.

The 96 real ones are corpus work rather than tool work and are part of the tier the entry sits in.

[tamnd]

Both boxes done. Tool side #57, content side tamnd/python-docs-vi-machine-translation#54. L02 139 to 86.

Two things this issue had wrong, both found by measuring rather than by reading.

Four of the six were already in the glossary. sys, builtins, exec and NaN were keep_en rows in v1. What was missing was not the rows, it was L02 reading them. import is the only row the content PR adds.

type and Boolean were on the wrong side of the triage here. This issue counted their 16 entries as real findings, and the glossary already held both as kept, so the exemption took them too. The glossary wins because it is the written decision. Whether it is the right decision is #58, which also covers object and module having no row at all while list and type have one, and all five being the same kind of Sphinx index category.

Infinity was proposed and then dropped. It had vô cực in v1 and the proposal flipped it to kept. Measuring first is what stopped that: the msgid is exactly Infinity in 1 entry, and "infinity" appears inside prose meaning the mathematical concept in 80. A row cannot say both, the prose wins, and that one entry stays an L02 finding.

The other direction is visible now too. G03 went 81 to 94: 13 entries where import appears in prose and the translation rendered it in Vietnamese anyway. The corpus was saying both things and nothing was reporting it.

The 96 real ones this issue set aside are now 86, and they are corpus work: module 41, object 36, Lists 3, 3 copyright notices, a Monty Python sketch, an Issue Tracking link and the Infinity entry above.