Update assistants to Fable

[hanna-paasivirta]

Short Description

Moves the main chat services (job_chat, workflow_chat, global_chat planner + doc_agent_chat prototype) from Sonnet to Claude Fable 5. Keeps RAG helpers, vocab_mapper, and the test judge on Sonnet.

This will increase our model costs, possibly by 5x. It is difficult to evaluate without running a lot of tests because thinking behaviour (i.e. output token volume) will vary between models and calls.

I haven't tested the effects of this update extensively. I skimmed over acceptance tests for the three chat services and tested locally with Lightning to see everything still works and doesn't seem slower. We'll be ready to rollback if there's complaints.

Fixes #524

Implementation details

• models.py: add "claude-fable" alias and CLAUDE_FABLE constant for claude-fable-5
• Point service configs at claude-fable: rag.yaml (model), gen_project_config.yaml, doc_agent_chat/config.yaml, global_chat/config.yaml (planner). Code fallback defaults updated to match.
• Keep on Sonnet: job_chat RAG calls (llm_search_decision, llm_retrieval), vocab_mapper.
• Triple max_tokens on Fable routes to absorb the new tokenizer (~30% more tokens): 16384 → 49152 (job_chat, workflow_chat, doc_agent_chat), 8192 → 24576 (planner).
• Pass an explicit per-request timeout=httpx.Timeout(600.0, connect=5.0) on the four non-streaming messages.create calls (job_chat, workflow_chat, doc_agent_chat, planner). Required: the SDK rejects non-streaming requests with max_tokens > ~21k unless a timeout is given. Values match the SDK default, so no behaviour change.
• Planner effort switched from high to medium.
• Remove dead temperature config: unread keys in doc_agent_chat, workflow_chat, and planner configs, plus the planner's unused self.temperature. Live temperature=0 settings (RAG, vocab_mapper, router) are unchanged and stay on Sonnet/Haiku, which accept it.

AI Usage

Please disclose how you've used AI in this work (it's cool, we just want to know!):

• Code generation (copilot but not intellisense)
• Learning or fact checking
• Strategy / design
• Optimisation / refactoring
• Translation / spellchecking / doc gen
• Other
• I have not used AI

You can read more details in our Responsible AI Policy

[hanna-paasivirta]

@josephjclark Does billing need to be adjusted for users?

[josephjclark]

Running the acceptance tests and it's certainly working!

Very hard to get a handle on better or worse. I think I'll release and smoke test staging.

We'll have to look back in a couple of weeks and assess whether the extra cost is worthwhile