[huntharo]

Stacked on #114. Only the last commit (ssh: treat remote paths as POSIX paths on every client) is new; the earlier four disappear from this diff once #114 merges.

Problem

SSH hosts are Unix: the remote command, the embedded scanners and the delete script are POSIX sh and Linux tools. But the client joined, split and compared remote paths with std::path, which follows the client's rules. On a Windows client:

• path_of joined /srv/data and logs into /srv/data\logs. Marks, the trail, the agent prompt and the exported delete script then named files that don't exist on the host. With root /, realpath decides whether the script deletes a file literally named \logs.
• A Unix name containing \ or : (common on Linux) was rejected, which failed the whole scan.
• PathBuf equality on Windows treats \ as a separator, so marks on /a/b\c and /a/b/c collided.
• ~ shortening, breadcrumbs, the trail separator and the indexing line all split remote paths by Windows rules.

Change

• disktree_core::path_style::PathStyle: Native is std::path, used for local scans. Posix is /-separated, used for SSH hosts. It provides join, parent, file name, ancestors, relative names, names and is-below. The app picks one from remote.is_some(). On Unix clients the two agree, so local and Unix-client behaviour is unchanged.
• Through PathStyle: tree::path_of, marks::find/refresh/is_hidden/display_path, crumbs_for_path, breadcrumbs and crumb labels, mark containment (toggle_mark, marked_ancestor, the remote plan filter, marks_ancestor), short_name, the trail separator, and the indexing progress line.
• Marks are indexed by exact path text, not PathBuf equality.
• Protocol: a host reporting platform: "windows" is refused by name, instead of being half-read with the wrong separator. The header's home and disk_root are now checked as POSIX-absolute, like root. The Windows-only rejection of \ and : in names is gone, because those names are now joined correctly.
• CLI: disktree ssh HOST PATH refuses a PATH that doesn't start with /.
• AGENTS.md: the invariant and the "where changes belong" table name PathStyle.

Tests

• path_style unit tests for both styles: \ and : in names, name boundaries (/srv vs /srvx), the root, trailing slashes, and C:\/\x not being POSIX-absolute.
• Hand-built Unix-host streams in remote.rs that run on every client. One checks that names with \ and : arrive whole. The other checks that a Windows host, and a non-POSIX home or volume, are refused.
• A window-harness test that loads a Unix host's tree with logs\old/a:b, checks crumbs, the mark, the trail and ~ display, and copies the delete script. On a Windows runner this is the case that used to break.
• Four round-trip tests that run the helper on the test machine are now cfg(unix), because that machine is the host. (Fix the macOS and Windows CI failures on main #114 already made the live preview dedup test Unix-only.)
• For the same reason, the three disktree-scan executable tests that read a snapshot or a discovery result are cfg(unix). The disconnect test reads neither and still runs everywhere.
• Window tests that pretend a local fixture is an SSH session now root it at a / path (remote_view_over): the temp directory on Unix, /srv/fixture on Windows.

Verified locally on macOS: cargo xtask lint and cargo xtask test pass, and cargo clippy -p disktree-core --all-targets --target x86_64-pc-windows-msvc -- -D warnings is clean. The app crate can't be cross-checked for Windows from macOS, so CI is the first run of these tests on a Windows client.

Note for local runs: gpui-omarchy's theme watcher watches $HOME. While other processes write there, app tests can abort with "Detected activity on thread notify-rs fsevents loop". Running with HOME set to an empty directory avoids it. This is unrelated to this change.