feat(voice): conversation-guided elevation + spoken interface configuration (ADR-110 D3b)

Voice inside the immersive interface becomes the primary elevation guide and
a control surface for the interface itself — fully local (Whisper STT in,
Kokoro TTS out), per ADR-110 D3b.

- src/actors/elevation_voice.rs: pure voice-signal layer — normalised n-gram
  ConceptIndex over elevatable labels (frontier stubs + working pages),
  greedy longest-match mention harvesting, VoiceDemandLedger with 30-minute
  half-life decay and excerpt/speaker provenance, explicit-intent parsing
  ('elevate X', 'formalise X', 'make X a class')
- elevation_actor: transcription stream subscription; conversational demand
  is now the PRIMARY candidate ranking (degree breaks ties / carries the
  queue when nobody is talking); voice-driven cases open at high priority
  with mention counts and utterance excerpts in the review surface; explicit
  spoken commands jump the queue and are confirmed aloud via Kokoro;
  concept index rebuilt from each graph snapshot
- src/actors/voice_interface_actor.rs: spoken view/graph configuration
  requests route to the SAME settings assistant the Control Center command
  box drives (settings_assistant_task extracted from bots_handler for
  reuse); conservative verb+noun intent gate keeps ordinary speech out;
  spoken confirmations; active whenever the local speech stack is up
- per-user/room attribution slots in when the XR LiveKit voice path lands —
  the ledger already models speakers

20 module tests green (7 ACSP contract, 11 elevation+voice, 2 interface
intent); clippy clean.

Co-Authored-By: jjohare <github@thedreamlab.uk>

Author: jjohare

docs/adr/ADR-110-agentic-actors-acsp-control-surfaces.md

@@ -90,6 +90,32 @@ infrastructure; actors compose it rather than re-implementing transport.
 queued candidates: sync governance (force-resync action), physics health,
 agent telemetry.
 
+### D3b — Voice is a first-class guide (local Whisper in, local Kokoro out)
+
+Conversation inside the immersive interface guides elevation **primarily**,
+and drives interface configuration. Both directions stay sovereign: STT is
+the local Whisper provider, confirmations speak back through local Kokoro
+TTS. Two consumers tap `SpeechService::subscribe_to_transcriptions`:
+
+- **Elevation demand** (`src/actors/elevation_voice.rs`): transcription lines
+  are matched against a normalised n-gram index of the graph's elevatable
+  labels (frontier stubs + working pages, rebuilt per cycle). Mentions feed a
+  decaying demand ledger (30-minute half-life) that becomes the *primary*
+  candidate ranking — degree only breaks ties or carries the queue when
+  nobody is talking. Voice-driven cases carry provenance (mention counts,
+  utterance excerpts, speakers when the voice path attributes them) and open
+  at `high` priority. Explicit commands — "elevate X", "formalise X", "make X
+  a class" — jump the queue entirely and are confirmed aloud. Per-user/room
+  attribution slots in when the XR voice path (LiveKit) lands; the ledger
+  already models speakers.
+- **Interface configuration** (`src/actors/voice_interface_actor.rs`): spoken
+  view/graph requests ("hide the ontology nodes", "increase spring strength")
+  route to the **same settings assistant** the Control Center command box
+  drives (`settings_assistant_task` → agentbox LLM → settings REST), with a
+  conservative verb+noun intent gate so ordinary conversation is never
+  hijacked. One assistant, two mouths: typed in the UX control centre or
+  spoken in the immersive session.
+
 ### D4 — Elevation, the flagship case
 
 `ElevationActor` (env-gated: `ELEVATION_ACTOR_ENABLED=1`):