docs: align Gemma 4 provisioning guidance with hotfix #275

[bram673a3a1a6f1170a8c6b148ab-ctrl]

Why this needs follow-through

#275 fixed the blank-spinner / 30-minute timeout failure, but the user-facing docs still do not explain the changed behaviour or how to distinguish normal provisioning from a failure.

Also, please use the shipped name consistently: this is Gemma 4 E2B Q4_0 on the llama.cpp lane, not “Gemma 3”.

Current, verified behaviour

• The default local model is Gemma 4 E2B Q4_0, downloaded from google/gemma-4-E2B-it-qat-q4_0-gguf (about 3.1 GB).
• On a cold ClawBox Connect, provisioning may compile llama.cpp with CUDA and download the GGUF. It is therefore variable and may run for much longer than a few minutes.
• The systemd ceiling is now 2 hours. It is a safety limit, not an expected wait time.
• The Local AI wizard now starts provisioning asynchronously, shows live journal-tail detail, and reports a systemd failure as soon as the unit enters failed; it should no longer leave the user at a silent spinner.
• The in-app updater re-delivers systemd unit files, so an already-installed box must complete System Update before retrying local AI.

Documentation changes requested

Add a short “Local AI / Gemma 4 provisioning” section to Choose Your AI Provider, linked from AI Providers, with this user guidance:

Before retrying: install the latest update from Settings → System Update, let the box reboot, then return to Settings → AI Provider → Local AI and enable/retry llama.cpp.

What you should see: a changing detail line while ClawBox prepares llama.cpp and downloads Gemma 4. The first install time depends on connection speed and whether the local runtime needs building; do not promise a fixed “couple of minutes.”

When it has failed: an explicit error/detail is a failed provisioning, not normal progress. Retry only after the update has completed. For support, include the displayed error and, if using SSH, the last journal lines:

journalctl -u 'clawbox-root-update@llamacpp_install.service' -n 100 --no-pager

Update the local-model wording in ClawBox Connect and Hardware Requirements so it does not imply 1–4B is the only realistic local option while the product ships Gemma 4 E2B Q4_0 as the recommended llama.cpp model.

Acceptance criteria

• No public/user-facing reference calls this hotfix or its shipped model “Gemma 3”.
• No user guidance gives a fixed short completion promise.
• Existing-device users are told to update before retrying.
• Users are told that changing live detail is normal; a surfaced failure detail is actionable.
• The support command matches the actual provisioning unit.