0.4.2 — 2026-04-03

EDC16 hardware-number extraction fix, evidence-based detection system, confidence scoring reframed as identification quality, and documentation consistency pass.

Fixed — EDC16 Extractor

• Hardware number extraction — _resolve_hardware_number() was too narrow (only searched the calibration area for 0281xxxxxx). BMW EDC16C31/C35 and some VAG PD bins store the HW number in the boot sector or active header and were missed entirely. The method now searches five regions in priority order: active header → calibration area → extended active window → boot sector → full-binary fallback. Supports spaced/dotted variants (0.281.xxx.xxx) with normalisation.
• OEM part number extraction — new _resolve_oem_part_number() method extracts vehicle-manufacturer part numbers (VAG format 03G906016J, BMW format 12 14 7 626 350). Previously hardcoded to None.
• ecu_family vs ecu_variant classify bug — ecu_family was set to the variant string (e.g. EDC16C8) instead of the canonical EDC16, causing --organize to create a folder per variant instead of a single scanned/Bosch/EDC16/ directory. Now ecu_family is always "EDC16"; the specific sub-variant is stored only in ecu_variant.

Added — Evidence-Based Detection

• Evidence tag system in BaseManufacturerExtractor — 16 standard evidence-tag constants across three categories: structural (SIZE_MATCH, MAGIC_MATCH, HEADER_MATCH, POINTER_TABLE, LAYOUT_FINGERPRINT, BOOT_BLOCK), identity (FAMILY_STRING, DETECTION_SIGNATURE, IDENT_BLOCK, FAMILY_ANCHOR), and cross-check (EXCLUSION_CLEAR, FILL_PATTERN, SYNC_MARKER, MANUFACTURER_CONFIRM).
• DetectionResult dataclass — rich return type wrapping matched: bool and evidence: Tuple[str, ...] for future migration of can_handle().
• Instance plumbing — _last_evidence attribute, _set_evidence() helper, and last_detection_evidence property on the base class.
• detection_evidence field added to the identity dict in identifier.py — the service reads evidence tags after can_handle() and passes them through to the result.
• Scan CLI now injects detection_evidence and detection_strength from the extractor instance into the identity dict before scoring, ensuring evidence-based scoring works in batch mode.

Changed — Confidence Scoring

• Reframed as identification quality — the score measures how reliably the system detected and extracted identity fields, not whether the binary content is unmodified. A tuned file with intact ident blocks scores identically to a stock file. All docstrings, CLI help text, and documentation updated to reflect this.
• DetectionStrength enum deprecated — retained as a fallback for extractors not yet upgraded to evidence tags; marked with .. deprecated:: docstring.
• No scoring weights or tier thresholds were changed.

Changed — Documentation

• confidence.md rewritten around the identification framing; tier descriptions, signal explanations, and filename-penalty rationale all updated.
• about.md — added "Learn more" cross-reference section linking to CLI ref, recipe format, confidence, contributing, and disclaimer.
• commands/identify.md — tier descriptions aligned with new confidence framing; added "See also" section.
• commands/cook.md — validate strict → validate before; added "See also" links.
• commands/families.md — CONTRIBUTING.md plain text → proper markdown link.
• recipe-format.md — added "See also" section and back-navigation link.
• bosch-internals.md — tier thresholds updated (Medium 25–54, Low 0–24, Suspicious < 0); added link to confidence.md.

Tests

• 4,751 tests passing (up from 4,734 in 0.4.1).
• 18 new EDC16 test methods across two new test classes: TestExtractHardwareNumberExpanded (11 tests — active header, boot region, mirror region, petrol format, spaced/dotted normalisation, region priority, 2 MB extended window) and TestExtractOemPartNumber (6 tests — absent OEM, key presence, VAG detection, suffix letters, space normalisation, type checking).