v1.3.0

Performance

Major performance improvements to pattern generation, achieving up to 40x speedup on warm queries and matching or exceeding kensaku.vim (jsmigemo + denops IPC) performance.

Warm queries (dictionary caches populated)

Input	v1.2.0	v1.3.0	Speedup
"a"	130ms	3ms	43x
"jo"	24ms	1ms	24x
"joutai"	0.6ms	0.03ms	20x

Key optimizations

• Eliminate coroutine overhead — Replace coroutine.wrap/yield with callback-based iteration, enabling LuaJIT JIT compilation of hot rank/select loops
• Iterative AA-tree insertion — Convert recursive insert_node/add_to_tree to iterative loops, enabling LuaJIT JIT compilation (previously caused NYI trace aborts)
• Multi-level caching:
  • Parent cache (FFI array) for LOUDSTrie reverse_lookup
  • Reverse lookup result cache per trie node
  • Mapping result cache per dictionary node (critical for incremental search)
  • Flat search result cache per query key
  • Query result memoization (repeated queries return in ~0.002ms)
• Pre-build parent cache at dictionary load time, shifting ~20ms from first query to load time
• Inline UTF-8 decoding in TernaryRegexGenerator, avoiding table allocation per word
• utf8_char result caching for repeated Unicode-to-UTF8 encoding

Compatibility

• jsmigemo pattern compatibility — Fix predictive romaji conversion to not recursively expand sokuon (っ). With the same dictionary, luamigemo now produces identical patterns to jsmigemo (only difference is kensaku.vim's \m prefix).
  • Before: じょう\%([たちつてと]\|っ[たちつてと]\)
  • After: じょう[たちっつてと] (matches jsmigemo)

Benchmarks

Added benchmark and comparison scripts in bench/:

• profile.lua — Internal phase-level profiling
• practical_bench.lua — Diverse query benchmark (80+ queries, incremental typing simulation)
• compare_patterns.lua — Pattern output comparison vs kensaku
• find_differences.lua — Comprehensive pattern difference scanner

See bench/README.md for usage.

API

No breaking changes. All public APIs remain backward compatible.