Copy code entries instead of translating them

[tamnd]

P07 says a doctest or a literal block is byte-identical to its msgid, because these are copied and never translated. 31 entries in the corpus were not, and the two ways they got there are both a layer being more forgiving than the one that checks it.

26 came in as human

sync.human_segments took any translated non-fuzzy entry from the mirror, and 136 of those are code. human is a provenance and not a grade: it says a person typed the string, and 30 of the 136 are a person having typed over the code. From tutorial/introduction.po:

File "<stdin>", line 1, in <module>     the English
File "1", line 1, in 2                  the translation

<stdin> and <module> are gone. Elsewhere it is the indentation inside a for body flattened to one space, the carets under a syntax error no longer under what they point at, and a column-aligned option table reflowed to single spaces. Every one is an example a reader copies out and then has to debug.

What they were made for is the comment translation, and that is worth less than the code. Comments are M8, with a prompt of their own and a check that every code line came back byte-identical, which is what classify.is_doctest has said all along.

The other 106 are already byte-identical and lose nothing by going. apply mints them from the msgid with passthrough=doctest on them, which is the same string with an accurate account of where it came from instead of a claim that somebody translated it.

5 came in as machine

From a run made before the classifier could recognise them. batch filters non-prose out, so nothing queues one now, but the segments outlived the rule that let them be asked. python fibo.py <arguments> was written into the corpus as python fibo.py <đối số>.

apply now copies a code entry whatever the memory holds, so the guarantee does not depend on no stale segment existing. The classifier has got stricter twice and will again, and each time it does the memory acquires another handful of translations of things that turned out to be code. What a string is worth is decided by what it is now, not by what an earlier run thought when it asked.

What is not touched

The human guard still comes first, so neither clause can quietly replace a reviewed string with the English. --refuzzy is the only way past it, which is how the 136 already in the corpus were rebuilt, and it was measured first: it touches those 136 and nothing else.

Only code is dropped, not everything the classifier calls non-prose. One no-op in the corpus is a :ref: whose display text a person translated correctly, and that is a bug in is_noop rather than a licence to throw the translation away.

load_human reconciles

It used to extend. An extend leaves all 136 in the memory for apply to write back, and the alternative was editing them out of the manifest by hand, which would make the memory something the content repo can no longer be rebuilt from.

Dropping is safe because the mirror is the only place a human segment comes from. There is no command that promotes a string to human, deliberately, so anything of that provenance was read out of Transifex and can be read again. Machine segments are left alone, being the one thing here that cannot be rebuilt without spending the run.

Measured

Over the real corpus, not a fixture.

	before	after
P07 code entries copied exactly	31	0
P05 format specifiers match	1	0
hard checks failing	6	4
findings	1,492	1,459
checks passed	28	30

P05's single finding was inside one of the same doctests: % operator read as a % o specifier, and the translation turned it into % t.

apply --check is clean. make check is green at 1,376 tests and 97.45% coverage.

Refs #11.