Docs: the audit, the coverage, and the real-data checks behind them

[Sherin-SEF-AI]

Documentation for #28 to #33. Every number is from a run on this branch's code, not typed.

Updated

• J1939: 26 PGNs, 152 parameters by the J1939-71 layouts, 30 more named, the preferred address table; the audit that found the old table's wrong bytes and wrong messages; the truck-log cross-checks confirming the new one.
• NMEA 2000: every standard PGN canboat defines (216), hand-written decoders first; canboat credited.
• OBD-II: 78 PIDs; DTCs via modes 03 and 07 with UDS 0x19 as fallback; a silent vehicle reported as silent.
• ISO-TP: tested against can-isotp; the STmin defect that found.
• openpilot logs: rlog, qlog, .bz2, .zst with the bundled cereal schema; canlab[openpilot].
• Calibrator: checked on two real cars from comma.ai, stated precisely per drive.
• Counts: pytest -q 753; real-data checks 99 (36 + 63, plus one private phase); stress run 33.
• tests/real_data/README.md: the J1939 cross-checks, phase_reference and phase_openpilot.

Limitations, updated honestly

• One physical adapter has now been used: a CANalyst-II on a Tata Tigor EV (111,006 frames). It has no listen-only mode. Others only on the virtual backend and stand-ins.
• UDS, ISO-TP, security access and OBD-II: scripted responders and independent implementations, not a real ECU. OBD-II formulas checked against J1979's worked values; no real OBD-II capture.
• RTS/CTS: synthetic frames and two can-j1939 nodes; still no real recording.
• J1939 covers 26 groups; SAE sells the full list.

Removed

The site's install page and limitations offered a "Vision OCR" option for reading a value off dashboard video. No such code exists in the repository.

No em-dashes. Site regenerated with python docs/site_src/build_site.py.