feat: live-validated against Ollama Cloud; schema instruction fallback for cloud

Phase 8: all six live checks pass (tier aliases, thinking-tier chat, real
tool invocation, structured Generate[T], forced failover with bench+skip,
skill agent). Discovery: ollama.com ignores the format field — the
provider now also states the schema as a system instruction (constrained
decoding locally, instruction-guided JSON on cloud), with hermetic test.

Co-Authored-By: Claude Fable 5 <[email protected]>

provider/ollama/wire.go

@@ -138,6 +138,14 @@ func (m *model) buildRequest(req llm.Request, stream bool) (*chatRequest, error)
 			}
 		}
 	}
+	if len(req.Schema) > 0 {
+		// Belt and braces: local Ollama enforces the "format" schema by
+		// constrained decoding, but Ollama Cloud ignores the field
+		// (verified live 2026-06-10) — so the schema is also stated as an
+		// explicit instruction. Harmless where format works, essential
+		// where it doesn't.
+		sys = append(sys, "Respond with a single JSON object that validates against this JSON Schema — no markdown, no code fences, no prose before or after the JSON:\n"+string(req.Schema))
+	}
 	if len(sys) > 0 {
 		out.Messages = append(out.Messages, chatMessage{
 			Role: "system", Content: strings.Join(sys, "\n\n"),