feat(R03): unify model routing and centralise the provider catalogue

EPIC R03 (Team Duy) — Model Providers & Routing. All six tasks done.

R03-T02 — Provider catalogue
  domain/models/provider_descriptor.py     ProviderDescriptor (frozen), WireProtocol, AuthKind
  infrastructure/providers/provider_registry.py
                                           thread-safe registry: id/alias lookup, dynamic
                                           lookup by model id, adapter selection by protocol
  providers/factory.py                     drops its own _REGISTRY table and delegates to the
                                           registry, still raising ProviderError for callers

R03-T03 — RoutingApplicationService (pure Python, 4 modes)
  application/model_routing/routing_models.py
                                           RoutingMode (off/auto/manual/fallback),
                                           RoutingRequest (immutable snapshot), RouteEvaluation,
                                           RoutingOutcome
  application/model_routing/routing_application_service.py
                                           the single decision flow, reached through two narrow
                                           ports plus a caller-supplied confirm callback, so no
                                           Qt import is needed
  application/model_routing/core_routing_adapter.py
                                           binds the ports to core/routing and AppContext

  Fallback is a new resilience mode: keep the selected model while it can serve the turn,
  re-route only when it cannot. Wired end to end through config.py, state.py,
  ui/routing_toggle.py and i18n.py (EN/JA/VI).

R03-T04 / T05 — Remove the duplicated routing flow
  ui/chat_panel.py (#L638), ui/co4e_tab.py, ui/folder_tab.py each drop ~35 lines of copied
  logic and call the shared service; the widgets now only build a RoutingRequest, host the
  Manual-mode modal and render the outcome.

R03-T06 — Token usage as an event
  infrastructure/telemetry/usage_sink.py   UsageEvent + UsageEventSink protocol, with tracker,
                                           in-memory and composite sinks
  providers/openai_compat.py, providers/anthropic.py
                                           publish a UsageEvent instead of writing to the
                                           usage tracker themselves
  core/usage_tracker.py                    adds current_context() so a sink can borrow and
                                           restore a thread's attribution

R03-T01 — Contract tests
  tests/contracts/test_providers.py parametrises over every provider in the registry: chat()
  signature, canonical assistant message, normalised tool calls, response closed, tool schema
  translation, ProviderError, list_models/test_connection, one UsageEvent per turn.

Test infrastructure fix (required to verify any of the above): tests/conftest.py used to put
the repository's PARENT directory on sys.path, so `import cowork_local.*` resolved against
whichever sibling folder happened to carry that name — on a dev machine, an unrelated older
checkout. The suite reported green while exercising different code. The conftest now binds
this checkout to the cowork_local name in sys.modules.

Verification
  pytest tests/                    236 passed in ~1.8s (102 before this change)
  scripts/check_imports.py         PASS, 0 forbidden imports in domain/ and application/
  new production files             largest is 288 lines, all under the 400 LOC ceiling
  new tests                        134 (50 contract, 70 unit, 14 integration), all offline

scripts/run_quality_gate.py does not exist yet (R10-T02), so DoD item 7 was covered by
check_imports.py plus the full suite.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

providers/anthropic.py

@@ -292,19 +292,31 @@ class AnthropicProvider(Provider):
                 args = {"_raw": b["json"]}
             tool_calls.append({"id": b["id"], "name": b["name"], "arguments": args})
 
-        # Dashboard usage event — real counts from the stream's usage events,
-        # else a ~4 chars/token estimate. Never breaks the turn.
+        # Usage event — real counts from the stream's usage events, else a
+        # ~4 chars/token estimate. Published to the telemetry sink (R03-T06)
+        # rather than written straight to the Dashboard store, so the provider
+        # stays a pure transport adapter. Never breaks the turn.
         try:
-            from ..core import usage_tracker as ut
+            from ..infrastructure.telemetry import usage_sink
 
             if usage_seen:
-                ut.record(self.name, self.model, usage_seen.get("in", 0),
-                          usage_seen.get("out", 0), usage_seen.get("cache", 0))
+                usage_sink.publish(usage_sink.UsageEvent(
+                    provider=self.name,
+                    model=self.model,
+                    input_tokens=usage_seen.get("in", 0),
+                    output_tokens=usage_seen.get("out", 0),
+                    cached_tokens=usage_seen.get("cache", 0),
+                ))
             else:
                 sent = json.dumps(payload.get("messages", []), ensure_ascii=False)
                 got = "".join(text_parts) + "".join(b["json"] for b in blocks.values())
-                ut.record(self.name, self.model, ut.estimate_tokens(sent),
-                          ut.estimate_tokens(got), 0, estimated=True)
+                usage_sink.publish(usage_sink.UsageEvent(
+                    provider=self.name,
+                    model=self.model,
+                    input_tokens=usage_sink.estimate_tokens(sent),
+                    output_tokens=usage_sink.estimate_tokens(got),
+                    estimated=True,
+                ))
         except Exception:  # noqa: BLE001
             pass