[flyinprogrammer]

Closes #303

Summary

The NL stem boost splits file stems and directory names with split_identifier, which returns ["__git"] for __git, so the keyword git never matches. Strip leading and trailing underscores first, in the stem boost only:

 _boost_stem_matches(boosted, query, max_score)
   for each candidate file
-    parts = split_identifier(path.stem)
-    parts += split_identifier(path.parent.name)
+    parts = _split_path_part(path.stem)          # split_identifier(name.strip("_") or name)
+    parts += _split_path_part(path.parent.name)
     boost by the share of query keywords matching parts

split_identifier and BM25 tokenization are unchanged.

Evidence

The new parametrized cases in test_apply_query_boost_nl_stem_match_boosts cover a leading-underscore case, ("git status", "lua/telescope/__git.lua"), and a trailing-underscore case, ("git status", "lua/telescope/git__.lua").

• Before (main's boosting.py): FAILED test_apply_query_boost_nl_stem_match_boosts[git status-lua/telescope/__git.lua]. The score stays at 0.5. The trailing case already passes through the prefix match and is kept as a regression guard.
• After: 4 passed.
• Full suite: make test 393 passed; make lint and make typecheck clean.
• Benchmark: run_benchmark NDCG@10 goes from 0.852 to 0.854. Few files are named this way, so the effect is small.

Merge Danger

Door: two-way

A ranking-only change with no index or cache format change; reverting restores the old ranking exactly.

Blast Radius: ranking

It only affects natural-language queries over files or directories whose names start or end with _.