Let the glossary decide which single words L02 leaves alone

[tamnd]

Closes #49, the tool half of it. The glossary half is content tamnd/python-docs-vi-machine-translation#54.

Narrowing the identifier rule in #47 put 6,558 single-word entries into Corpus.prose() for the first time and took L02 from 10 findings to 144. A third of those were entries reading sys, builtins, import, exec, NaN and Infinity. Each one is a Sphinx index entry naming a module or a statement, and a reviewer left it in English because that is what a Vietnamese programmer calls it.

Nothing in the string separates those from module, object and type, which are the other side of the same list: ordinary English words used as index categories, sitting untranslated. sys and Notes are the same shape, and that is precisely the discrimination #47 established the classifier cannot make.

So the decision is made once, by hand, in the glossary, where keep_en already means it. L02 now skips an entry whose whole msgid is a kept term, and G03 checks that term in both directions instead. A written decision rather than an exception buried in a check.

Matched on the whole msgid, not on a substring: a kept term inside a sentence says nothing about whether the sentence was translated.

Measured on the real corpus

check	before	after
L02 no entry is the English verbatim	139	85
G03 kept terms stayed in English	81	95
G02 glossary renderings are used	191	190
G04 no untranslated glossary term	143	142

54 fewer L02 findings, not the 38 the issue predicted, and the extra 16 are worth naming. The glossary already held type and Boolean as kept, and the issue's triage had counted those 16 entries as real findings. The glossary and the triage disagree, the glossary wins because it is the written decision, and the disagreement is now visible instead of being an argument nobody was having. Filed as #58.

The 14 new G03 findings are the other direction: entries where a kept term appears in prose and the translation rendered it in Vietnamese anyway. sys is hệ thống in some entries and sys in others, so the corpus says both, and G03 is now the place that says so.

Checks

make check green, 1,403 tests, 97.45 per cent coverage. H03 clean.

[tamnd]

Correcting the table in the description. Those numbers were measured with an Infinity row in the glossary proposal, and that row was dropped before the content side shipped: vô cực is right for the 80 prose entries that mean the mathematical concept and wrong for the 1 index entry, and a row cannot say both.

Measured on what actually shipped, content tamnd/python-docs-vi-machine-translation#54:

check	before	after
L02 no entry is the English verbatim	139	86
G03 kept terms stayed in English	81	94
G02 glossary renderings are used	191	191
G04 no untranslated glossary term	143	143

53 fewer L02 findings rather than 54, and the single Infinity entry stays reported. G02 and G04 do not move at all, which the earlier table had wrong.

The rest of the description holds. type and Boolean are exempt because the glossary already kept them, the issue's triage had counted them as real, and that disagreement is #58.