m45-native-short-key-memory-attribution

M45 Native Short-Key Latency And Memory Attribution Evidence

M45 closes as a partial native-engine result with measured blockers. It does not claim browser, frontend, WASM, public-demo, packaging, deployment, or product-delivery speed.

Verdict

• Short-key latency: partial. hao passes the <=3.0x same-run librime gate, but n and ni still miss and remain measured benchmark-parity blockers.
• Memory: resident target only. Track A steady after-ready resident samples are below 107,797,708 B, but the first startup still reaches a real 127,475,712 B peak. M45 records steady-state-meets-target-standing-peak-cost and does not claim full memory success.
• Storage and behavior: Track A keeps rsmarisa_byte_backed, mmap table/prism mapping, zero selected heap mirrors, source_fallback=false, positive rsmarisa counters, page-bounded output/context, and short-key candidate output matching upstream librime 1.17.0.

Evidence Map

• Phase 0 native baseline: phase-0-native-baseline/
• Phase 0 short-key candidate oracle: phase-0-short-key-oracle/
• Final native benchmark: final-native-benchmark/
• Final short-key candidate comparison: final-candidate-comparison/
• Phase 0 verdict: phase-0-verdict.md
• Final memory attribution: final-memory-attribution.md
• Visual evidence: visuals/

Final quality gates are recorded in final-native-benchmark/final-gates.md.