0.6.3 — 2026-08-11

Shared-axis detection, scoring refinements, speed-based deduplication of the Rust/Python backends, and musl wheels for Alpine Docker.

Added — Map Scanner Improvements

• Shared-axis detection — after a valid [X][Y][data] table is found, the scanner probes forward for additional blocks of identical geometry sharing the same X and Y axes. Catches the common ECU pattern where fuel, timing, boost, and EGR tables sit consecutively after one pair of RPM×Load breakpoints.
• --show-series — group tables sharing identical X/Y axes with indented └─ continuation rows in human-readable output.
• --max-series-tables — new CLI parameter (default 16, 1 = off).
• Axis value plausibility bonus — axes whose values look like memory addresses (tiny span relative to baseline, e.g. [16798, 16804, 16810, …]) now get a ×0.25–0.60 quality penalty. Eliminates high-scoring false positives from code-pointer tables and encoded data structures.
• Sentinel penalty — tables with >25% placeholder values (0x0000, 0xFFFF, 0x8000, 0x7FFF) now receive a ×0.40–0.80 multiplier. Pushes partially-erased flash blocks and sentinel-heavy coincidental structures toward the bottom of the ranking.

Changed — Scoring & Defaults

• Default --min-score raised from 0.75 to 0.85. Combined with the two new penalties above, the default filter now returns ~70% fewer tables while keeping

  90% of genuine calibration maps.

• Version string always shows (rust) — see Removed below.

Fixed

• max_results=None no longer crashes scan_map_tables. Pass None for unlimited results (the Rust function now treats 0 as "no cap").

Removed — Dual Backend Elimination

Every CPU-bound function was benchmarked. For each function, only the fastest implementation survives. No more dual maintenance, no more fallback paths, no more oracle fuzz tests.

• shannon_entropy / is_low_entropy — Rust only (36–75× faster). Python _py_* reference implementations deleted.
• count_unique_in_window / find_unique_context — Python only (CPython's C-level bytes.find is 1.1–1.6× faster than Rust's memchr loop for ECU context-anchor workloads). Rust implementations deleted.
• find_changed_blocks — Rust only (102× faster). Python reference deleted.
• scan_map_axes — Rust only (115× faster — 139 ms vs 16 s). Python reference deleted.
• scan_map_tables — Rust only (24× faster — 2.5 s vs 60 s). All Python scoring, pairing, and dedup logic deleted. map_hunter.py shrinks from 1,433 to 244 lines.
• Oracle fuzz tests removed (470 tests). No dual implementations to compare against — each function exists in exactly one language.
• OPENREMAP_FORCE_PYTHON env var has no effect. The Rust extension is mandatory.
• Dead Rust code removed — count_unique_in_window and find_unique_context from entropy.rs (153 lines).

Added — CI/CD

• musl wheels (musllinux_1_2_x86_64) — cross-compiled in CI for Alpine Docker images. pip install openremap on Alpine automatically selects the musl wheel instead of the glibc one.