[thekevinbot]

Summary

Speed-of-light e2e test for recursive globs: './**/*.md' over a deep, wide tree, and './a/b/**/*.md', where a/b holds a quarter of the tree.

• The tree mixes in .txt files, dotfiles and dot-named directories, which globstar with dotglob off must not return.
• Native is find, pruning dot-named entries. It beats bash globstar here.

Fixes #1306

Status: dirsql misses the bar

Rows match native for both globs. Speed does not.

Test run on main (startup 0.04s):

Glob	native	dirsql	ratio
'./**/*.md'	2.149s	7.182s	3.3x
'./a/b/**/*.md'	0.975s	6.380s	6.5x

Direct timing on a 1.6M-file tree built by the test's own builder (release CLI, warm cache):

Query	find	main	with #1319 and #1321
./**/*.md	0.55s	2.09s	1.43s
count(*) over ./**/*.md		1.21s	0.80s
./a/b/**/*.md	0.21-0.37s	0.56s	0.35s

• The prefixed glob already walks only a/b.
• The remaining gap has the same causes as Glob speed: one-level listings ('./*', './docs', './') #1305: the walk itself, then SQLite stepping and JSON output.
• The leaf directories hold about 800 entries each, so the per-directory parallel judging from Speed up the path-table walk on large directories #1321 never engages. A parallel walk across subdirectories is part of Path-table walk: readdir-level speed for the count(*) floor #1322.

Fixes land in PRs stacked on this one.

E2E Verification

N/A: test-only change. The new e2e test is the red. No receipt, because no package source changed.