Split an archive entry the way the format defines, not the host

[otsobide]

Windows CI went red on dev after #87 merged (test (rust, windows), two failures in apps/core/tests/names.rs). The PR run was green because the cross-OS matrix only runs on shipping branches, which is the gate working one merge too late.

Chasing it turned up three problems with one cause, plus a fourth that predates all of this.

One cause

normal_components split entry names with Path::new(name).components(). std::path is #[cfg]-dependent while NameRules is data, so the rules were portable and the splitting they ran over was not.

Unix regression, currently on dev. A backslash is a legal Unix file name character and was not a separator, so dir\f.txt was one component that rewrite then refused. Collapse could build an archive it would not extract:

collapse compress tree -o tree.zip   exit 0, and it passes the #86 verification
collapse extract tree.zip -o back    exit 1, `back` is never created
   error: the archive entry "tree/a\b.txt" cannot be written:
          "a\b.txt" becomes "a\b.txt", which is not a name this system can hold

All three formats, and it fails at extract time, which is after the user may have deleted the originals. In the desktop it was worse than an error: nameProblem renders only inside the naming sheet, which is not open on that path, so the Extract button did nothing and said nothing.

Windows, issue #63's own hole. A leading a: parses as a drive prefix, and a prefix is not a Normal component, so it was silently discarded. a:b/c.txt was judged, reported and written as b/c.txt: the report never asked about that colon, on the only platform the colon matters.

The CI failure itself. rewrite_entry("a:b/CON/c?.txt.") answered b\CON_\c_.txt.

An archive entry name is not a host path. ZIP mandates / (APPNOTE 4.4.17.1) and tar has used it since v7, so entry_components splits on / everywhere and NameRules::windows() now answers the same from a Mac as from Windows. The backslash moves into WINDOWS_REJECTED, where it belongs: an ordinary character that Unix holds and Windows cannot.

That also removes the post-rewrite contains('\\') check, which was dead code except when it was wrong. check_replacement already refuses both separators before either is pushed, so it could not fire for its stated reason.

The fourth, which is older

zip and 7z sanitized an entry name, joined it to the output and wrote, with no check on where the join landed. tar has always resolved the parent through unpack_in. Measured on dev before this branch, with a symlink named link already in the output directory:

extract(archive holding "link/evil.txt") -> Ok(["link/evil.txt"])
file written outside the output directory: true

origin/main has the same shape, so this is in the released v0.7.0, not something #87 introduced. We never create symlinks, but extracting into a directory that already has one is ordinary.

ensure_inside now resolves the directory about to be written into at every zip and 7z write site. It also covers the Windows case where PathBuf::push clears the buffer on a component that parses as a drive (docs/c:evil.txt resolving against the current directory of C:), and sanitize_entry_path re-parses each component to close that at the source.

Tests

585 Rust (was 578) and 113 Vitest (was 112). Every new test was checked against the unfixed code rather than assumed:

test	fails without the fix
an_entry_splits_the_same_way_on_every_host	yes
only_windows_refuses_a_backslash_inside_a_component	yes
a_unix_name_holding_a_backslash_survives_the_round_trip	yes
the_report_sees_a_colon_in_the_first_component_too	Windows only, pin
no_format_writes_through_a_symlink_already_in_the_output	yes
a_renamed_entry_cannot_be_written_through_such_a_symlink	yes
a_colon_in_a_later_component_cannot_clear_the_path_being_built	Windows only, pin
desktop: naming refusal with no dialog open	yes

The two assertions that failed on Windows CI compared the message against the path the test built, while the message is rendered from a canonicalized root: on Windows that carries a \\?\ verbatim prefix and expands the runner's 8.3 short name. They now compare against what extraction actually resolved. The identical assertion in apps/cli/tests/names.rs was a third Windows failure that run never reached, because make stops at the first red app.

Notes

• Labelled full-matrix, since the point is the Windows and macOS legs.
• The Windows run died inside apps/core, so remote, cli and server-backend have still never been tested on Windows. This branch is the first run that can get past core.
• Closes the docs half of core: an entry name with a colon passes the traversal guard and becomes an NTFS stream on Windows #63 (docs/threat_model.md had nothing on entry names).