feat: browser access (allowed_origins, CORS with Private Network Access), job cancel, Tomie/Unity adapters, benchmark

Includes earlier uncommitted work: Tomie and Unity adapters, benchmark, structured output, call metrics.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

local_translator/engine.py

@@ -10,7 +10,7 @@ from .provider import Provider, ProviderError
 from .storage import encode
 
 log = logging.getLogger(__name__)
-PROMPT_VERSION = 2
+PROMPT_VERSION = 3
 
 
 class Engine:
@@ -78,7 +78,7 @@ class Engine:
                     self.storage.item_result(identity, item["position"], "completed", cached, cached=True)
                 else:
                     pending.append(item)
-            while pending:
+            while pending and self.storage.job_state(identity) != "cancelled":
                 batch, size = [], 0
                 while pending and len(batch) < options["batch_size"]:
                     item = pending[0]
@@ -112,10 +112,13 @@ class Engine:
             for item in remaining:
                 item["attempts"] += 1
             error = "Model omitted the item or returned a duplicate ID"
+            call_error = None
             try:
                 output = await provider.translate(model, project, [{"id": item["id"], "text": protected[item["id"]].text} for item in remaining])
             except ProviderError as exc:
                 output, error = {}, str(exc)
+                call_error = str(exc)
+            self.storage.record_call(identity, len(remaining), getattr(provider, "last_metrics", {}), call_error)
             failed = []
             for item in remaining:
                 item_error = error