feat(inference+mobile): sovereign inference on our mesh + installable mobile PWA (twin or box) - #31

… mobile PWA (twin or box)

Two asks, one capstone: run LLMs on OUR infrastructure (not a cloud provider), and see/reach it
from a phone.

Sovereign inference (tools/inference.py):
  - a model is an immutable DATA SPHERE (provenance, pinned integrity, residency-fenced); loading
    weights needs a read Grant; a mutated model is un-citable.
  - inference_service_workload: serving a model is a GPU workload the compute plane places on a
    TRUSTED backend (Needs firewall keeps a sensitive model off untrusted/volunteer/vendor).
  - route_inference: FAIL-CLOSED sovereign-first. A sensitive prompt (or a residency-fenced model)
    goes to a sovereign endpoint or BLOCKS — it never leaves for a cloud LLM. Non-sensitive may fall
    back to a vendor connector only when policy allows.
  This is the difference between "our infrastructure" and "a cloud provider like Claude."

Mobile PWA (portal_server.py):
  - installable (manifest.webmanifest + service worker, offline-ish shell), mobile-responsive, apple
    web-app tags; a twin/box endpoint badge (SOURCEOS_ENDPOINT).
  - /api/inference surfaces the sovereign posture + per-model routing; a "Sovereign inference" console
    section shows sensitive prompts routing sovereign or blocked.
  Access model: default to the TWIN (always-on rendezvous — the box sleeps, the twin doesn't; grants
  + coordinator live there), opt-in DIRECT to the box for LAN/offline. Same reference-vs-direct
  lattice from the mount analysis, applied to control access.

capd/sovereign-inference.mesh.capd.json. Tests: +7 inference +4 portal = 156 tools tests green.