perf(sftp): size the tar decision from listings, not a stat per file (#494)

[kipavy]

Closes #494.

When the exec size probe (find … -size / Get-ChildItem) can't answer (no probe for the shell, a failure, or the 120 s timeout), has_large_remote walked the tree itself to find a file ≥ 64 MiB. That walk duplicated walk()'s traversal and stat'ed every entry: one round trip per file plus two per directory. SFTP listings already carry each entry's type and size.

The issue offered "go per-file directly". I didn't do that: it would send every small-file tree on a probe-less host (including Windows OpenSSH's default cmd.exe) per file instead of tar, which is a slowdown. This keeps every routing decision as it was and makes the fallback cheaper.

Change

• resume/mod.rs: walk()'s traversal becomes visit(), which takes a callback returning Go / Prune / Stop. walk() is rebuilt on it with the same output: same order, same symlink skipping, the same unsafe-name skipping, and the same report_skipped for unreadable metadata.
• resume/large.rs: the fallback is now large_by_walk() on visit(). On SFTP it costs one round trip per directory, none per file, and stops at the first large file. For a tree of 10,000 files in 200 folders, that's about 200 round trips instead of about 10,400.
• Listed.complete: whether the listing vouches for the entry's type and size. large_by_walk stats anything it doesn't vouch for, exactly as before:
  • SFTP: complete only when the server sent permissions, plus a size for non-directories (listed_completely).
  • Local disk: always complete.
  • Docker, FTP, WebDAV (From<RemoteFile>): never complete, since e.g. docker's listing writes stat || echo 0 and can't tell an unknown size from 0. Those endpoints keep the old per-entry lookups.
  • Unreadable metadata (stat: None): also stat'ed individually, as before.

Tests

• New: a large file is found at any depth; only the selected items are stat'ed and each folder is listed once ((1, 5) for a 4-folder tree); the walk stops at the first large file ((1, 1)); a listing that doesn't vouch is stat'ed entry by entry and still finds the large file; listed_completely for each attribute gap.
• TestFs counts stats and lists, and can serve bare listings.
• Full cargo test --lib against this PR's base (69ddb1bd), in the same container: base 789 passed, this branch 794 passed (the 5 new tests). Both fail the same 3 commands::plugins tests, because the JS plugin bundles weren't built in that container. cargo fmt --check and cargo clippy --all-targets -D warnings are clean.

Not measured

• A timing run against a real server. The speed-up is shown by the stat and list counts in the tests.

🤖 Generated with Claude Code