voice-limit-guard: live revoke never fired — LiveKit's JSON uses snake_case permissions

livekit-server (verified on 1.13.7) serialises ParticipantPermission as
can_publish / can_publish_sources; the reconciler read canPublish /
canPublishSources, saw 'publishes nothing' for everyone and never called
UpdateParticipant, so turning Allow Screen Sharing off did not stop an
in-progress share. Normalise the permission keys before deciding.
Verified on a local Synapse + LiveKit + guard stack: share track gone
from the SFU 2 s after the policy flip. Tests added (44 pass).

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PPmy3tPq869XDW4njjVaKA

livekit/voice-limit-guard.py

@@ -399,10 +399,27 @@ def reconcile_publish_sources(current, forbidden: set):
     return sorted(effective - forbidden)
 
 
+def _camel(key: str) -> str:
+    head, *rest = key.split("_")
+    return head + "".join(part.capitalize() for part in rest)
+
+
+def normalize_permission(perm: dict) -> dict:
+    """LiveKit's Twirp JSON serialises ParticipantPermission with proto field
+    names (`can_publish`, `can_publish_sources`, ...) — verified against
+    livekit-server 1.13 — while the JWT grant and older docs use camelCase.
+    Return a camelCase copy so the policy code reads one shape. (protojson
+    accepts either spelling on input, so the copy we send back is fine.)"""
+    out = {}
+    for key, value in (perm or {}).items():
+        out[_camel(key) if "_" in key else key] = value
+    return out
+
+
 def reconcile_participant(alias: str, participant: dict, forbidden: set) -> bool:
     """Enforce the forbidden-source policy on one live participant. Returns True
     if an UpdateParticipant call was issued."""
-    perm = participant.get("permission") or {}
+    perm = normalize_permission(participant.get("permission") or {})
     if not perm.get("canPublish", False):
         return False  # publishes nothing -> nothing to revoke
     current = perm.get("canPublishSources") or []