fix(partial-matcher): let ** match zero path segments - #215

Fixes #187.

The bug

getPartialMatcher is the predicate that decides whether a directory is worth descending into. It walked the pattern parts and the input parts in lockstep with a single index, which meant ** could only ever stand for one or more segments — as soon as the ** matcher rejected the current input part, the whole pattern was abandoned.

** also matches zero segments, and that case matters as soon as the segment right after the ** is the one that matches. With dot: false, picomatch's ** does not match .foo, so:

await glob('**/.foo/*.ts', { cwd, expandDirectories: false });
// tinyglobby: []
// fast-glob:  ['.foo/foo.ts']

The crawler never entered .foo, so the correct full matcher downstream never got the chance to run. With dot: true the ** matcher accepts .foo, which is why the issue only reproduces on the default.

The same shape breaks any pattern where a ** precedes a segment the ** itself won't match — **/.deep/a/**/*.txt returned [] too.

The fix

Track the pattern position (j) and the input position (k) separately. When a ** doesn't consume the current input segment, advance only the pattern and retry the next pattern part against the same segment, instead of ending the walk. When it does consume it, the existing early return true is unchanged.

The two indices only diverge in that one case, so the common path costs the same. noglobstar still treats ** as a literal part.

Validation

• pnpm test → 135 pass, 0 fail. Reverting only src/utils.ts fails all 5 new assertions, so they pin the fix rather than restate current behavior.

• pnpm typecheck, pnpm check (biome), pnpm build all clean.

• pnpm bench on the typescript-eslint fixture — no regression, tinyglobby still fastest in both:

  	tinyglobby	fast-glob	glob
  packages/*/tsconfig.json	2627 ops/s	2598	2249
  **/*	50 ops/s	35	19

Tests added: three unit tests on getPartialMatcher (leading **, mid-pattern **, and a negative case that the looser walk doesn't turn into a false positive) plus two integration tests against the existing .a / .deep fixtures.

I did not touch the dot semantics themselves — this is purely the traversal predicate being stricter than the matcher it guards. The other half of #187 (whether dot: false should skip dot directories at all, as opposed to dot files) is a deliberate behavior choice and is left alone.

🤖 Written with assistance from Claude Code.