telescope-fzf-native (C)

A faithful C port of fzf's matcher — byte-identical constants and slab model — whose value to this survey is as the only in-process proxy for fzf's algorithm in cross-matcher benchmarks.

Language	C (Lua FFI wrapper for Neovim/telescope)
License	MIT
Repository	nvim-telescope/telescope-fzf-native.nvim
Surveyed revision	b25b749b (all file/line citations pin this commit)
Category	Matcher library (no UI of its own)
Algorithm class	Smith-Waterman variant, no substitution, single matrix (fzf V2)

Overview

What it solves

telescope.nvim needed fzf's ranking inside the editor process — fzf itself is only usable as a child process. src/fzf.c (1272 lines) ports the whole matcher surface: fzf_fuzzy_match_v1, fzf_fuzzy_match_v2, fzf_exact_match_naive, fzf_prefix_match, fzf_suffix_match, fzf_equal_match, plus fzf_parse_pattern with the full !/^/$/' term syntax.

Design philosophy

Fidelity. The constants are byte-identical including the derived forms:

typedef enum {
  ScoreMatch = 16, ScoreGapStart = -3, ScoreGapExtention = -1,
  BonusBoundary = ScoreMatch / 2,
  BonusNonWord = ScoreMatch / 2,
  BonusCamel123 = BonusBoundary + ScoreGapExtention,
  BonusConsecutive = -(ScoreGapStart + ScoreGapExtention),
  BonusFirstCharMultiplier = 2,
} score_t;

and the slab allocator is faithful (fzf_make_default_slab() = {100*1024, 2048}), including the V2→V1 degradation on slab overflow.

The divergences — all Unicode, all TODO-marked

static char_class char_class_of(char ch) {
  return char_class_of_ascii(ch);
  // if (ch <= 0x7f) { return char_class_of_ascii(ch); }
  // return char_class_of_non_ascii(ch);
}

Every non-ASCII byte classifies as CharNonWord, producing spurious boundary bonuses inside multi-byte codepoints, and normalize_rune is an identity no-op. Scores diverge from real fzf on any non-ASCII input. For benchmarking ASCII path corpora it remains the best available in-process stand-in for fzf's algorithm — which is how the comparison's independent numbers use it.

Analysis spine (delta-only)

This port intentionally has no design of its own; against the fzf deep-dive's spine it differs only in: Unicode & case handling (ASCII-only classification, no normalization — above) and incremental & streaming architecture (none; telescope drives it per-keystroke over Lua FFI, re-matching the full list). Algorithm, scoring, prefiltering, and memory strategy are the upstream design verbatim.

Key design decisions and trade-offs

Decision	Rationale	Trade-off
Port verbatim, constants included	Ranking parity with fzf is the product	Inherits the single-matrix non-optimality
Skip the Unicode tier (TODOs)	ASCII covers most file paths	Non-ASCII scores are wrong, not merely unnormalized
C + Lua FFI	In-process, editor-embeddable	Manual memory management across the FFI boundary

Sources

• src/fzf.c — the port (constants, matchers, slab, pattern parser; the char_class_of TODO quoted above).