0.7.1 — unreleased

0.7.x stabilization — diff-maps robustness + layout consumers + ISSUE-2.

Added

Same-file-only tier (ISSUE-2) — cook --allow-non-unique now stamps recipes with non-unique anchors as metadata.portability = "same_file_only". tune refuses to apply such a recipe to any binary whose sha256 differs from the source (ecu.sha256), unless the new --force flag is passed (loud warning; Phase 1/2/3 mechanical checks still run and can still abort). validate before reports the mismatch (same_file_only block; safe_to_patch forced false). Unblocks ~25% of real tunes whose edits sit in non-unique regions, without weakening the portability promise: strict default unchanged, the exception is machine-enforced.
diff-maps near-match pre-index — _near_match_pass indexes remaining stock tables by shape, turning the worst-case O(unmatched × remaining) into O(unmatched × same-shape candidates) — a tune that rescales many axes no longer blows up quadratically (audit follow-up item 3).
cook / cook-volatile region tags — every recipe instruction now carries a region field (calibration / code / erased / mixed / unknown) from the flash-layout segmenter; edits outside a calibration region get a CODE_AREA flag and a prominent portability warning ("may not apply to other revisions of this ECU"). Advisory only: never filters, never blocks, tune ignores the tags. No calibration signal → unknown, nothing flagged. The layout scan is shared with --annotate-maps (one scan for both map annotation and region tags) — default cook adds ~0.2 s on a 4 MB pair; --no-annotate-maps adds the full scan (~2.5–3 s).
scan-maps / diff-maps calibration-region default — both commands now scan the detected calibration region by default instead of the whole file: junk tables from code / erased sectors are hidden and counted (layout_filtered / tables_hidden in JSON, a note in human output). --whole-file restores the full scan; --region overrides; no calibration signal falls back to whole-file behaviour automatically. A wrong layout estimate can only change what the report shows — never recipe or patch output.
diff-maps correlation near-match — tables whose axis breakpoints changed between stock and tuned are now paired up instead of silently dropping into the only-in-* lists: same shape + close axes (edited breakpoints, within ~15% of the axis range) + strongly correlating cell grids (Pearson r ≥ 0.95) → reported as a match flagged ↺ axes changed, with the old/new axis values in JSON (axis_stock / axis_tuned).
Correlation on every match — each match now carries a correlation value in JSON and drives a refined ⚠ suspicious flag: near-total cell change is suspicious only when the grids do NOT correlate (low r = two different maps sharing axes); a heavily retuned map (all cells changed, r ≈ 1.0) is recognised as a legitimate retune.
"Changed but not identified" report — changed byte regions no matched table covers and that are not table-shaped are listed explicitly (human section + unidentified_changed_count / unidentified_changed in JSON + Markdown export section). Every changed byte in the binary is now accounted for: matched map, promoted table, or listed here.

Changed

attach_maps() / tag_instruction_regions() accept a precomputed table scan; cook / cook-volatile share one stock scan between map annotation and region tags instead of scanning twice (default cook on a 4 MB pair: 8.05 s → 5.70 s).

Fixed

diff-maps no longer silently misses an entire rewritten map when the tuner re-scaled its axis breakpoints (previously documented as "intentional" — now reported as an axis-changed near-match).
diff-maps strided near-match mis-diff — compound (interleaved) tables with rescaled axes were diffed with contiguous cell reads, mixing the two halves (audit 2026-08-24: 0x23C0A8 reported 25/25 changed vs correct 20/25). The near-match shape guard now compares stride and strided tables are read with their stride, mirroring the exact-match path.
diff-maps "changed but not identified" is now per-byte: blocks straddling a matched table's edge report their uncovered sub-ranges (previously dropped by any-overlap); matched spans carry the ±4 pad alignment slack so a table's own alignment drift never produces false unidentified tails.