Harden current-information search failures

Preserve required-search semantics for live hints and expose stable gateway error codes without leaking provider details.

docs/openapi.yaml

@@ -480,8 +480,10 @@ paths:
         and output-budget policy from `capability` plus optional `taskKind`. It
         never infers task type from `input` or `context`, and clients cannot
         supply provider parameters. Ordinary AI questions allow model-selected
-        server-side web search; current-information questions require it. Other
-        tools remain disabled. Search has no separate credit fee; settlement
+        server-side web search and retain thinking if search is unavailable.
+        Current-information questions require search and never degrade to an
+        unverified offline answer. Other tools remain disabled. Search has no
+        separate credit fee; settlement
         uses the provider-reported LLM input and output Token counts, including
         any search context charged by the provider.
         For account grants, `oobe` is accepted only for dictation polish and the
@@ -1785,6 +1787,13 @@ components:
           type: array
           maxItems: 20
           items: { type: string, minLength: 1, maxLength: 64 }
+        taskKind:
+          type: string
+          enum: [ai_question, current_information_question]
+          default: ai_question
+          description: |
+            AI execution intent for this card. Current-information cards require
+            server-side web search; ordinary cards match hold-to-talk AI policy.
         metadata:
           type: object
           additionalProperties: true