[HayronHgh]

Python callers previously had to cross a process/JSON boundary to use FastK. This publishes the current 0.2 native NumPy interface, its core ownership/IO fixes, and a fixed backtest that changes only the storage adapter.

The backtest uses identical RSI(14) 30/70, SMA20/60 and 60-bar momentum strategies, next-open fills and carried monthly state. All five stores deliver the same readonly int64 OHLCV layout. The annual BTC/ETH experiment has 180 isolated application runs; every equity curve, trade digest, fee and final position matches. Relative to Arrow, monthly total medians improve by 2.61%, 8.80% and 0.79%; the last is effectively tied within observed variation.

A separate Q1 cohort rebuilds public commit 814c566 unchanged and compares its actual CLI/JSON path against native delivery: 30 complete backtests with identical outcomes. These are public-delivery/application ratios, not Rust-only algorithm speedups. The repository has no formal 1.0 tag; the initial commit is version 0.1.0-rc.1.

This branch consolidates the previously local native work, bounded positional reads, lock/recovery and snapshot contracts, lifecycle/cache controls, and their immutable evidence. It retains historical negative results and the initial Arrow import-boundary mistake, which was corrected before all five formal rounds were rerun. The adaptive read channel remains a design proposal.

Validation on Windows GNU:

• 142 Rust tests and one doc test passed on the current source.
• Seven backtest correctness tests passed, covering causal fills, costs, batching, gaps, failure cleanup and no replay.
• Rustfmt and Python Ruff checks passed. CI now runs the backtest correctness tests in the native matrix.
• Annual 130-query preflight and Q1 16-query preflight match canonical packed bytes; all 210 formal application outcomes agree within their own cohorts.
• Evidence/source/archive hashes verified; isolated version-study data and metadata remain unchanged.

Native Linux/MSVC execution is not claimed by these local measurements. OS caches are uncontrolled and warmed by preflight. Data preparation, process imports and resident-control loading boundaries are documented. Raw measured files are kept byte-for-byte under docs/evidence/.

Start with docs/BACKTEST_RESULTS.md, docs/VERSION_COMPARISON.md and external-tests/backtest/README.md.