fix(ext/node): normalize recursive fs mkdir dot-segment paths

[nathanwhitbot]

Summary

• Normalize paths passed to Deno's filesystem API for recursive mkdir so lexical dot-segments like a trailing /. do not make recursive creation fail (matches mkdir -p). The fix lives in RealFs::mkdir.
• Add focused coverage for sync and callback recursive node:fs mkdir paths containing literal dot segments, plus Deno.mkdir unit coverage.

Closes #35450

Scope

Because the fix lives in RealFs::mkdir, it changes behavior for both node:fs mkdir/mkdirSync and Deno.mkdir/Deno.mkdirSync — e.g. Deno.mkdir("foo/.", { recursive: true }) now succeeds instead of failing. This is intentional (covered by tests/unit/mkdir_test.ts) and is a reasonable correctness improvement matching mkdir -p, though it is broader than the ext/node title suggests.

.. segments are deliberately preserved: Path::components() strips . but keeps ParentDir, so a/../b/c/. → a/../b/c and the filesystem resolves .. itself, avoiding lexical symlink-flattening issues. Pinned by the "preserves parent segment" test.

Tests

• env -u RUSTC_WRAPPER cargo test -p unit_node_tests --test unit_node -- _fs::_fs_mkdir_test
• ./target/debug/deno fmt --check tests/unit_node/_fs/_fs_mkdir_test.ts
• git diff --check

Regression coverage was also checked by temporarily reverting the implementation, rebuilding deno, and confirming the new dot-segment tests fail with ENOENT before restoring the fix.

[nathanwhitbot]

Pushed fix(node): preserve mkdir parent segments to fix the Node compat test-fs-mkdir.js failure. Recursive mkdir now only strips trailing /. segments instead of resolving the whole path, so a/../b/c keeps Node’s lexical parent creation behavior.

[nathanwhitbot]

Pushed a CI follow-up: replaced the direct path.slice() call with StringPrototypeSlice for prefer-primordials, and adjusted the new mkdir dot-segment test so the lexical parent assertion only runs on non-Windows while still checking the requested target path.

[nathanwhitbot]

Pushed a formatter-only follow-up for the recursive mkdir path normalization helper. The failing lint job was dprint asking for the long return expression in stripTrailingCurrentDirSegments() to be split over multiple lines.

[nathanwhit]

are you sure this shouldn't be fixed in the rust part of the impl instead of a js normalization? this affect Deno.mkdir as well...

[nathanwhitbot]

Good catch. I moved the recursive mkdir path normalization into the Rust ext/fs implementation so it covers Deno.mkdir/mkdirSync too, and left Node’s wrapper using the original path for return/error semantics. Added Deno namespace coverage for recursive /. paths.

[nathanwhitbot]

Pushed fix(node): remove unused fs primordial import to address the failing lint jobs. The CI failure was no-unused-vars for StringPrototypeSlice in ext/node/polyfills/fs.ts, and the new commit removes that stale import.

[bartlomieju]

Nice, small targeted fix. A few notes:

Drop the ext/node/polyfills/fs.ts hunks — they're pure formatting churn. Both changes only reflow Deno.mkdir(path, { recursive, mode }) from one line onto three; there's no functional change. The PR description says these "preserve original validated paths," but that preservation actually comes for free from not normalizing on the JS side — it isn't implemented by these hunks. Removing them keeps the PR minimal.

Scope note for the description: because the fix lives in RealFs::mkdir, this also changes Deno.mkdir/Deno.mkdirSync (e.g. Deno.mkdir("foo/.", { recursive: true }) now succeeds), not just node:fs. That looks intentional (you added tests/unit/mkdir_test.ts coverage) and is a reasonable correctness improvement matching mkdir -p, but the title says ext/node while the behavior change is broader — worth calling out.

Add a one-line comment on normalize_recursive_mkdir_path noting it works around DirBuilder::create failing on trailing dot-segments (link #35450). It reads as generic normalization otherwise.

Otherwise LGTM. Good call preserving .. — Path::components() strips . but keeps ParentDir, so a/../b/c/. → a/../b/c and the filesystem still resolves .. itself, avoiding lexical symlink-flattening issues. The "preserves parent segment" test pins that nicely.

[nathanwhitbot]

Thanks! Addressed all three:

• Reverted the ext/node/polyfills/fs.ts reflow hunks — fs.ts is now byte-identical to main.
• Added a comment on normalize_recursive_mkdir_path noting it works around DirBuilder::create failing on trailing dot-segments, linking fs.mkdirSync({ recursive: true }) fails for paths containing dot-segments (e.g. "./a/b/c/./") #35450.
• Reworked the PR description with an explicit scope note that the fix in RealFs::mkdir also changes Deno.mkdir/Deno.mkdirSync, not just node:fs.