[P0][LOCAL-01] Offline/local/hybrid inference policies

Programme: #1
Epic: #21

## Objective
Implement supervised local inference with enforceable offline/local-with-connectors/hybrid modes and no hidden cloud fallback.

## Scope
- Inference worker adapter with model load/unload, request/cancel, health and resource reporting.
- `Offline`: deny network for inference and use only installed/cached data.
- `Local inference + online connectors`: keep model local while separately permitting granted connector/web network.
- `Hybrid`: explicitly named remote providers with destination-scoped context grants.
- Failover policy that requests a mode change rather than silently escalating remotely.
- Per-request destination/telemetry trace.
- Runtime adapters initially compatible with llama.cpp-class local runtimes without tying product state to one backend.

## Security / product rules
- Local inference is not synonymous with anonymous browsing; network activity is reported separately.
- Remote provider permission is scoped to selected context/destinations.
- Model workers have no direct authority over tools.

## Acceptance criteria
- [ ] Offline mode technically prevents inference-worker network egress.
- [ ] Local-model failure cannot trigger remote inference without a new applicable grant/user-visible mode change.
- [ ] Connector networking remains independently controllable from inference networking.
- [ ] Cancellation works while generation/model loading is active.
- [ ] UI exposes active mode/provider/model and where inference runs.
- [ ] Network-denial and hidden-fallback regression tests run in CI.

## Dependencies
- CORE-03
- SEC-02
- SEC-04

**First phase:** P0
**Maturity target:** P1
**Owner:** local-ai-speech

No implementation files, entry points, or tests are named. Start by reading the CORE-03, SEC-02, and SEC-04 dependencies and map the inference worker, connector networking, and mode/grant boundaries. Done means the listed acceptance criteria are implemented, including cancellation, visible runtime state, and CI regression tests for network denial and hidden fallback.

Written by the indexing model from the issue text.