Send a message to the console AI assistant

Runs ONE assistant turn on the caller's conversation and answers when it is done. The model sees a fixed system prompt, the SERVER-SIDE history of this conversation and text -- nothing else from the client. It may only call read tools (the resource layer and the flavor / image / network lists, all read back as the caller), navigate (an allowlisted console screen of this project) and propose_instance, which writes a PROPOSAL and creates nothing: the user executes it with .../proposals/{proposalId}/actions/confirm. Tool output is passed to the model as data. At most 8 tool rounds per message; links and images in the answer are removed. Tokens are billed (assistant.input.ktoken / assistant.output.ktoken). Validation: READ on compute -> 403 (compute:read for tokens); unreachable scope, or a conversation that is not the caller's / not in this project / expired -> 404; text 1-2000 characters -> 422. Feature flag: 404 while the assistant service is disabled. Async & state: Synchronous (bounded by a turn time budget); appends to the conversation, writes tool-call audit rows, proposals and token ledger rows. Errors: 401; 403; 404; 409 CONFLICT with details.reason SUBSCRIPTION_FROZEN (paused subscription) or ASSISTANT_UNPRICED (the token price is not published -- an unpriced assistant is closed, not free); 422; 429 with details.reason ASSISTANT_BUDGET_EXCEEDED (the subscription's daily token cap) or the edge rate limit; 500; 503 with details.reason ASSISTANT_UNAVAILABLE (no model provider configured, or the provider failed). Idempotency/Concurrency: Not idempotent (each call is a billed model turn).