[thekevinbot]

Refs #1474, #1475, #1476

The walk built each entry's path with rel.join(name), a malloc plus a realloc per entry on every worker thread; the allocator's arena locks contended. It now allocates the exact length once.

Local, 400k files, load 6-9, same-machine A/B against origin/main (hyperfine, 25 runs, interleaved), output md5 identical:

Glob	Before	After
./**/{a,b}*.md	58.5 ms	41-44 ms
./**/[a-c]*.md	62-63 ms	46 ms
./**/*.md	82-83 ms	82-86 ms (no change)

No case crosses its FASTEST tolerance, so no bar flips.

Red run: 38200140376 (Rust CI; both child_path_* unit tests fail on capacity, 32 vs 25).

Changelog / Migrations

• Changelog fragment under packages/rust/changelog.d/
• Migration fragment with the same slug under packages/rust/migrations.d/

E2E Verification

• Ran e2e suites locally for every affected SDK
• Python SDK e2e: pass
• TypeScript SDK e2e: pass
• Rust core e2e: N/A beyond the above
• Receipts: packages/python/e2e-attestations/bot-speed-gap-childpath-1011.json, packages/ts/e2e-attestations/bot-speed-gap-childpath-1011.json
• Command(s) run: pytest tests/e2e/glob_speed_{depth_bound,one_level,recursive,symlink,matchers,dotfiles,gitignore}_test.py -x -q; pnpm test:e2e
• Result summary: all pass