fix(score): rank exact filename stem match above fuzzy filename (#722)

[gustav-fff]

Closes #722

Root cause

crates/fff-core/src/score.rs:757 — is_exact_filename requires main_needle_len == fname_len and a case-insensitive equality against the ENTIRE filename (including extension). A query that exactly matches the filename stem (e.g. lsp ↔ lsp.lua) falls through into the fuzzy_filename branch capped at ~16% of base_score. lsp.lua (2-char subsequence in an 7-char filename) and lsp/typos_lsp.lua (2-char subsequence in a 13-char filename) both land there, differ by only 2 points, and any frecency/git bonus on the longer file flips the order.

Fix

crates/fff-core/src/score.rs — add is_exact_stem: true when the needle equals the filename's stem (bytes before the LAST dot) and there is no additional dot beyond that. Awards a 30% base_score bonus, positioned between fuzzy_filename (~16%) and exact_filename (40%). Adds a new match_type = "exact_stem" and includes it in exact_match.

Steps to reproduce

Add this unit test to crates/fff-core/src/score.rs inside mod filename_bonus_tests (uses the existing make_files/search helpers already in that module) and run against origin/main:

#[test]
fn issue_722_repro() {
    let (files, arena) = make_files(&["lsp/typos_lsp.lua", "lsp.lua"]);
    let results = search(&files, "lsp", arena);
    for (path, s) in &results {
        eprintln!("{} total={} bonus={} type={}", path, s.total, s.filename_bonus, s.match_type);
    }
    assert_eq!(results[0].0, "lsp.lua");
}

Run:

cargo test --lib --manifest-path crates/fff-core/Cargo.toml issue_722_repro -- --nocapture

Expected (bug): lsp.lua total=56 bonus=8 type=fuzzy_filename and lsp/typos_lsp.lua total=54 bonus=6 type=fuzzy_filename. Both files score as generic fuzzy matches; the 2-point gap is trivially overturned by any frecency/git bonus on lsp/typos_lsp.lua, causing it to rank first as the reporter observed in the screenshot.

How verified

• New regression test test_exact_stem_beats_fuzzy_filename asserts lsp.lua ranks first with match_type == "exact_stem".
• Full cargo test --lib in crates/fff-core: 125 passed, 0 failed.
• make test — all Rust + Lua + bun tests pass (pre-existing tsc env issue in test-node unrelated to change).

Post-fix output for the repro:

lsp.lua           total=127 bonus=27 type=exact_stem
lsp/typos_lsp.lua total=54  bonus=6  type=fuzzy_filename

Automated triage via Gustav. Honk-Honk 🪿

[dmtrKovalenko]

@gustav-fff this is very inefficient, find a way to make this good enough so i can pick it up

[gustav-fff]

[triage-bot] DIRECTED: pushed 9d8f63f.

Consolidated exact-filename / exact-stem into one branch with cheap length gates first:

• Skip the arena filename read entirely when neither needle_len == fname_len (exact) nor needle_len + 2 <= fname_len (stem) holds — the vast majority of candidates.
• When the read is needed, do it once (was twice: previously write_file_name_from_arena was called for both is_exact_filename and again for is_exact_stem, and even the pre-existing exact branch redundantly did it whenever simd_filename_match wasn't Some(m.exact)).
• Reuse the pre-allocated fname_buf (as the surrounding code already does) rather than allocating per file.

Net effect: on non-.-containing directory paths and on any file where needle_len differs from both fname_len and fname_len - k (k >= 2), the extra work vs. pre-#722 is zero. On the plausible set, exactly one arena copy + one memcmp.

Tests: cargo test --lib --manifest-path crates/fff-core/Cargo.toml — 125 passed, 0 failed. Regression test test_exact_stem_beats_fuzzy_filename still passes.

Let me know if you want a different rank position (currently 30% between fuzzy_filename ~16% and exact_filename 40%) or a stricter stem definition (e.g. reject dot-prefixed filenames).

Honk-Honk 🪿

[gustav-fff]

[triage-bot] DIRECTED: pushed 9d8f63f — length-gate the exact-filename and exact-stem checks before touching the arena.

Before: hot loop did up to TWO write_file_name_from_arena calls per candidate (allocates + copies filename bytes) — one inside is_exact_filename, another inside is_exact_stem, even when neither could possibly match.

After: single fused check. Cheap length-based pre-check first — exact needs needle_len == fname_len, stem needs needle_len + 2 <= fname_len. If neither is plausible, we skip the arena read entirely. Otherwise the filename is materialized ONCE and both variants read from the same buffer.

fname_buf is already reused across iterations (declared at score.rs:706) so no per-file allocation on the hot path when the length gate does trigger.

All 125 fff-core unit tests pass, including the test_exact_stem_beats_fuzzy_filename regression.

If you want it more aggressive (e.g. inline the ascii-lowercase compare from the arena chunks directly without materializing at all), say the word and I'll do that pass — it's a bigger change touching simd_path.rs.

Honk-Honk 🪿