Make the suite pass on Windows

[jakejackson1]

Stacked on #40, which adds the windows job. Review that one first; this PR's diff is only its own change.

The job ran 658 tests on its first outing and returned 45 failures. One was the library. The other 44 were the suite and the checkout assuming POSIX — which is exactly what a platform nobody had ever tested on was going to produce, and why the job went in continue-on-error.

Library finding 1 — pathinfo() splits on \ under Windows

FileInfo::setNameWithExtension() and Filename::sanitizeNameWithExtension() split a client filename with pathinfo(). That treats \ as a path separator on Windows and as an ordinary character on POSIX, so the same name split two ways:

Client name	POSIX	Windows
a\b.txt	a-b.txt	b.txt
..\..\windows\win.ini	windows-win.ini	win.ini
a----t----t----.php\x00.png	a-t-t-php-x00.png	x00.png

Filename::rewriteCharacters() rewrites \ to -, so the rule both layers read from Filename is that a backslash stays in the name. Filename's own docblock says the two layers "must not disagree about what the rules are" — and a rule that depends on which platform is applying it is not one rule.

Filename::splitNameAndExtension() owns the split now, with / the only separator everywhere. It answers what pathinfo()'s PATHINFO_FILENAME and PATHINFO_EXTENSION answered in every other respect, and the new provider pins that case by case — the trailing slash basename() drops, the dotfile that is all extension and no name, .. and ..., repeated slashes. Nothing on POSIX changes; the existing 76 sanitizing data sets pass untouched, which is the evidence for that.

On Windows this keeps name content that was previously discarded, so it is noted in UPGRADE.md as well as CHANGELOG.md.

Library finding 2 — the default configuration could not store a file on Windows/7.3

reserveDestination() confirms its exclusive create by comparing fstat() against lstat(); that is what catches an x that followed a symlink. Windows before PHP 7.4 reports ino as 0 for both, so the comparison had nothing to compare — and read that as a mismatch. Every reservation was answered 'Destination is a symbolic link', and a reservation happens whenever $overwrite = false, which is the default. That platform could not store a single file.

The comment above the check said the comparison "degrades to same-drive and detects nothing" there. It detected everything. Corrected along with the code.

The comparison is now skipped where the inode is unavailable rather than absence being read as mismatch. A real file has a real inode, so POSIX gives up nothing, and the symlink protections in this class were already documented as not load-bearing on Windows — where a symlink needs a privilege an uploading process should not hold.

This relaxes a symlink check, and I would rather it were ratified than inherited. My reasoning is that the pre-fix state was not protecting anything on that platform — it was refusing everything. If refusing to run on Windows/7.3 is preferable to skipping the check there, that is a legitimate alternative and your call, not mine.

testAReservationIsNotRefusedWhereTheInodeIsUnavailable() stubs lstatEntry() with what that platform answers — no inode, and a dev of its own, so it is the missing inode that decides rather than a lucky match on the drive. Removing the guard reproduces the exact 'Destination is a symbolic link' failure; I checked both ways.

Library finding 3 — an orphaned placeholder on Windows

releaseReservation() decided whether the destination was a symlink by whether readlink() failed. That is not the same question on every platform: PHP's Windows readlink() answers a regular file with its own canonical path rather than failing, so the 0-byte placeholder took the symlink branch, found nothing matching the inode it had opened, and was never removed — holding the caller's name against every later upload of it.

is_link() asks directly. The stat cache is cleared first, since reserveDestination() has already lstat'd that path.

Both branches already had POSIX tests — testAReservationThatCannotBeConfirmedIsNotReportedAsASymlink for the plain file, testReservationThroughASymlinkLeavesNothingAtItsTarget for the link. They pass on Linux either way; the first is what caught this, on the platform where readlink() behaves that way.

The other 44

Path separators in assertions (~35). upload() returns $this->directory . $filename and the constructor ends the directory with DIRECTORY_SEPARATOR, where these tests build their working directory with /. Windows resolves both to the same file, so the storage was right and the string comparison was wrong. Four sites and the shared assertStoredAs() helper now go through a destinationOf() helper. assertFileExists() was always fine — PHP resolves either separator.

Line endings (5 sites, 2 tests). The callback tests echo PHP_EOL against an expected literal \n. They are about hook order, so the literal wins and the expectation stays readable.

A CRLF checkout (3). CatalogueTest parses i18n/upload.pot line by line with /^msgid "(.*)"$/m. Under a CRLF checkout the trailing \r makes "$ unmatchable, so no msgid was read and every assertion failed vacuously. .gitattributes pins that file to LF — which the i18n workflow needs regardless, since its guarantee is a byte comparison against a generator that writes LF.

A subprocess (1). FilenameTest builds a php -r script containing "a\nb". escapeshellarg() quotes with " on Windows and cannot escape one inside the argument, so the inner quotes were dropped and PHP read a\nb as a constant. The script carries no quote character of its own now.

What genuinely cannot run there (2, errors). NTFS refuses control characters in a filename outright, so touch() cannot create the colliding file that two of the collision-message data sets need. Split into their own @group posix test, since a data set cannot carry a group — the same treatment the fourteen symlink()/chmod()/umask() tests get in #40.

Verified

phpunit (full, --exclude-group posix, --exclude-group mbstring), lint, phpstan, check-syntax, base64-docs, i18n:pot, the package archive diff, and composer validate --strict all pass locally. 647 → 671 tests.

Both Windows jobs now pass — 654 tests and 1,284 assertions on 7.3 and on 8.5, identical to --exclude-group posix locally, so the grouping selects the same work on both platforms and what remains is real difference rather than selection drift. continue-on-error is removed and the job is a gate.

One correction

While measuring for this branch I found the test counts I quoted on #40 were taken against the wrong baseline. main is 628 tests, so #40 is 628 → 647 (the +19 was right; the absolute figures were the CJK-tests branch I had been standing on). #40's description is corrected.

🤖 Generated with Claude Code