fix: pluralize only the messages Symfony pluralizes

[66Ton99]

Follow-up to 39493ab, which chose the plural form of a constraint message by the
locale. A review of it turned up one behaviour bug, one BC break, and a gap in
the tests. This fixes all three, plus what a review of this branch turned up.

The convention reached constraints Symfony does not pluralize

getPluralCount() fell back to a naming convention — a minMessage is chosen by
a numeric min — for anything absent from PLURAL_COUNT_OPTIONS. That fallback
is meant for a constraint of an application's own, but it also answered for
Symfony's, whose pluralized messages the table already covers in full:

Constraint	Message	Convention returned	Symfony pluralizes it
Range(min: 5, max: 10)	minMessage / maxMessage	5 / 10	no
Count(divisibleBy: 3)	divisibleByMessage	3	no
Image(minRatio: 1.5)	minRatioMessage	1	no
Image(minWidth: 100)	minWidthMessage	100	no

Harmless while those translations carry no |, since a single-part message comes
back unchanged. One that does carry a | is cut at it:

uk, Range(min: 1), 'Значення має бути {{ limit }} або більше (див. А|Б).'
  → 'Значення має бути {{ limit }} або більше (див. А'
en, Range(min: 5), the same string
  → 'Б).'

Symfony's constraints are now answered from the table and from nothing else.
Across all of Symfony's own validators.*.xlf catalogues a | appears in
exactly 11 message ids, and the table has a row for each, so nothing of
Symfony's is lost by this. A constraint pluralized in a later Symfony release
needs a row before it is chosen from, and keeps every form until it gets one —
the behaviour the library had before the table existed, which is the safe side
of the trade.

A constraint that extends one of Symfony's is answered the same way. The
table is matched with instanceof, so AppCount extends Count already took
Count's rows; the guard beside it looked at the class itself, so the same
constraint then took the convention for every message the table does not name.
AppRange extends Range was still choosing a form for its minMessage by
Range::$min. Such a constraint inherits Symfony's options and is validated by
Symfony's validator, which calls setPlural() for the table's rows and nothing
else, so the whole ancestry is looked at now.

translateMessage() takes two arguments again

svaroh_js_form_validator.factory.class is the documented way to replace the
factory, and translateMessage($message, ?array $parameters = null) predates all
of this. A third parameter on it means PHP fatals on any subclass that overrides
the method with the signature it has always had — not a deprecation, a
Declaration ... must be compatible at class-declaration time. 39493ab is not in
any tag yet, so nothing shipped broken and only a dev-main checkout could have
hit it; this keeps the break from reaching a release.

The plural path moved to translatePluralMessage(), which decides the count and
then hands the message to translateMessage() with that count as the %count%
parameter — the call it was making directly is the call translateMessage()
makes. An override therefore sees every message: pluralized, not pluralized,
and the ones no form could be chosen for. Three tests cover it, and declaring
the test's LegacyFactory would itself be a fatal error if the signature ever
slipped again.

A message no form could be chosen for is reported

The catch (\InvalidArgumentException) was silent. Swallowing it is right — an
unfinished translation should not take a form down — but the only trace left was
a message still showing its | separators in the browser. The factory now logs
it through an optional PSR-3 logger, naming the catalogue entry to go and finish.
It is a warning, not a debug note: that message is what the form actually
shows, which is a defect in the rendering rather than a diagnostic detail.

psr/log joins require at ^3.0; it was already present transitively through
symfony/http-kernel. The range is deliberately narrow — the logger the tests
collect records with cannot be declared against psr/log 1, whose
LoggerInterface::log() leaves $message untyped, so narrowing it in a subclass
is a fatal error.

Tests for the table

Every row of PLURAL_COUNT_OPTIONS is load-bearing now that the convention no
longer stands behind it, and one row always was: filenameTooLongMessage is
chosen by filenameMaxLength, which no convention could derive. It had no test.
Neither did Choice, WordCount, or Count::exactMessage. They do now, along
with a constraint extending Range and one extending Count — the latter both
keeping the table's rows and losing the convention for the rest.

The helper the tests parse a constraint through also had to learn that a File
constraint is exported as a plain option list rather than as itself, which is why
a File test could not have been written against it before. It now shares one
validator between the factory and the form, as the container does.

The two tests that asserted on Symfony's own English wording bring a catalogue of
their own now, so a rewording upstream cannot fail them for a reason that has
nothing to do with choosing a form.

Verified locally on PHP 8.5

• composer test: 95 tests, 268 assertions, green (81 / 250 before).
• composer phpstan: no errors.
• composer validate --strict: valid.
• The test container compiles with the setLogger call, resolving to logger.
• Removing the namespace guard fails the three original cases and nothing else;
  narrowing it back to the class itself fails exactly the two new subclass cases;
  reaching past translateMessage() again fails exactly the new override test.
  Each test locks its fix rather than the code.

Not changed

prepareMessage() in the browser still splits on |, because the fallback still
needs it. Its two-form rule is now reachable only through a translation with too
few forms, which 3_23.md says.

🤖 Generated with Claude Code