Don't send the developer role to Ollama.

Ollama Cloud returns a 500 for any request that carries both a `developer` role
message and an image. The same request succeeds with `system`, and text-only
requests with `developer` are fine, so the two only conflict together.

pi selects the `developer` role whenever a model reports reasoning support, and
the ollama case hardcodes `reasoning: true` for every model, so every image sent
to a vision model like gemma4:31b-cloud failed. The case carried no `compat` at
all and therefore inherited pi's OpenAI-shaped defaults; Workers AI already
pins the same flag just above.

Pin the role rather than the `reasoning` capability, which is reported
unconditionally and would disable thinking for ollama models that do support it.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SetKoutv8UUpQn2MeV66RF

Author: rlapray

packages/workshop-backend/src/ai-models.ts

@@ -563,6 +563,9 @@ function getModelDirect(config: AiModelConfig, sessionAffinity?: string): ModelH
           input: ["text", "image"],
           cost: ZERO_COST,
           ...window,
+          // Ollama Cloud 500s when a request carries both a `developer` message and an image; pi
+          // picks that role because we report reasoning: true above, so pin the role, not the flag.
+          compat: { supportsDeveloperRole: false } satisfies OpenAICompletionsCompat,
         },
         ...(config.apiToken === ""
             ? { apiKey: "unused", headers: { Authorization: null } }