Native comparison driver notes

Run the same steps on the base and head in their separate checkouts:
1. PROOF_FX_BIN=<real fx 0.0.14 executable> python3 mobile-seed.py <fresh-output-directory>.
2. Start dev:server with T3CODE_DEV_SEED=0, T3CODE_AUTO_BOOTSTRAP_PROJECT_FROM_CWD=0, --home-dir <fresh-output-directory>/home, --host 0.0.0.0 and a distinct port. Start Expo Metro from apps/mobile on all interfaces with a distinct port.
3. Connect the pool proxy, lease iPhone 17 Pro, install with vp run mobile:install ios --session <session>. Open the Expo development client at the Metro URL with disableFab=1, disableOnboarding=1.
4. Pair using the repository test-t3-mobile pair-client.sh helper. Use a fresh app connection per side; remove the previous connection through Settings > Environments > Manage > Expand details > Remove environment.
5. Chats > New chat > demo. Open Model and reasoning settings, Default (grok-3-mini), Medium. Open Runtime. Observe the selected mode and available choices.
6. Take two screenshots at least 500 ms apart and compare their bytes. Capture once the pair is identical. Inspect before uploading.

Actual device: iPhone 17 Pro, iOS 27.2, light appearance, 402x874 capture. Both sides had one demo project, no threads, identical defaultModelSelection fx/default, defaultRuntimeMode full-access and fx 0.0.14. Native client fingerprint was identical across these provider-only commits. The base selected Supervised and offered two modes; the head selected Full access and offered three modes.
