T1: canary_diff.py + hosted-CI pytest suite

[josephnef]

Summary

Adds the host-side half of the T1 canary cross-validation oracle: a Python differ (tests/canary_diff.py) that compares kernel and devourer canary dumps with runtime-ephemeral register masking, plus a 13-test pytest suite covering the parser, mask logic, channel-aware capture-state filter, and exit-code contract.

A dedicated path-scoped GitHub Actions workflow runs the pytest suite on every PR that touches the differ or its tests — no libusb, no VM, runs in about a second.

Masking rules

Without masking, every diff drowns in genuine drift sources that aren't init bugs:

• MAC 0x040 / 0x550 / 0x560 — per-queue / TBTT / beacon-window counters that increment continuously.
• RF[A|B] 0x42 — thermal-meter sample register; reads vary with chip temperature.
• BB 0xc1c / 0xe1c bits 31:21 — TX BB-swing tx_scaling_table_jaguar index, updated by PowerTracking8812a (and the kernel's phydm watchdog) from the thermal-meter sample. Other bits of 0xc1c (AGC table select, static base bits) ARE diffed.
• BB 0xc20 / 0xe20 at 5G only — CCK regs that the long-lived kernel iface retains from prior 2.4G activity, while devourer captures from a fresh process per run. Suppressed only when --channel > 14; the same divergence at 2.4G is reported as real.

--strict disables all masking. --show-clean prints a CLEAN summary line even when there's nothing to report.

Exit codes

• 0 — no real init-drift divergence
• 1 — divergences found in non-ephemeral registers
• 2 — file parse error / register-set mismatch

Test plan

• python3 -m pytest tests/test_canary_diff.py -v — 13 / 13 passing locally
• Smoke-tested against real 8814 ch6 captures — surfaced 21 real divergences cleanly, masked counters / thermal noise without false positives
• Hosted CI green on the new Canary diff unit tests job