Gen7FW: I2C drivers for onboard temp sensor + TI voltage rail monitor (datasheet-verified via host harness)

[datacrystals]

Stacked on the simulator PR. New Gen7 hardware support that the .ioc had configured but the firmware never initialized:

• I2C4/I2C5 init — MX_I2C4_Init (PD12/PF15, VOLTAGE_MONITOR) and MX_I2C5_Init (PF0/PF1, ONBOARD_TEMP_SENSE), timing 0x60404E72 from the .ioc, with a comment that a CubeMX regen must not duplicate.
• OnboardTempSensor — TMP102/TMP1075/PCT2075-family pointer-register sensor: probe, 12-bit and 13-bit-EM decode, onb_temp shell command, telemetry/alert wiring.
• RailMonitor — auto-detects INA226 / INA228 / INA3221 via ID-register patterns over a capability table, CAL/SHUNT_CAL math, per-channel poll, rails shell command, i2cscan probe utility.

No hardware was available, so this went through an extra datasheet audit instead. Four review agents cross-checked the drivers against the actual TI/NXP/ST datasheets (SBOS547C, SLYS021A, SBOS576C, PCT2075 rev 10.1, DS13313) and caught real bugs before bench ever saw them:

• TMP1075 config POR 0x00FF would have falsely latched the EM bit and reported 2x the real temperature (fixed: skip EM unless positively ID'd as TMP102-generic).
• INA228 power register decoded with the 20-bit shift used for VBUS/CURRENT but it's full 24-bit (would have under-reported power 16x).
• INA226/INA3221 die-ID exact-match rejected legitimate revision-nibble variants (US-fab INA226 = 0x2261); now masked.
• Warm-boot hazards: INA228 CONFIG forced to ADCRANGE=0 by RMW, INA3221 CH_EN+MODE RMW, INA226 reads the OVF math-overflow flag at the rails instead of trusting a wrapped sample, and rejected probes still record raw IDs for diagnosis.

Host-side proof without hardware: Images/Gen7FW/MainProcessor/tests/host/ is a plain-g++ harness (no dependencies) with accuracy-modelled mock devices — TMP1075 (real-0x00FF POR), TMP102 (pointer-masked), INA226 (real CAL transfer function), INA228 (rev-1 ID, 24-bit power), INA3221 (channel-enable gating), plus exact bus-transaction-sequence assertions. 297/297 checks pass locally.

The firmware cross-build through the real pipeline (Images/Gen7FW/build-main-rte.sh foc_demo.json Release, arm-none-eabi) links clean; the .rte-gen7/generated mirror is updated to match.

When the boards show up on the bench: docs/I2C_HW_Bringup.md has the exact power-up + first-scope-check + i2cscan sequence with expected IDs.