fix(claude): derive context token budget from the Claude model in Claude Code mode

[herikwebb]

Problem

In Claude Code mode, ai_service_manager.chat_model reflects the user's most recent non-Claude provider selection and is None on a Claude-only setup. The attachment token budget in on_message fell through to the legacy 100-token floor (80 tokens after the 0.8 factor), which:

• silently skipped cell-output attachments (their estimate always exceeds 80 tokens), and
• dropped every attachment after the first via the remaining_token_budget <= 0 break.

Change

• New _resolve_context_token_limit helper: in Claude Code mode the budget comes from the configured Claude model via model_info_from_id (200K-window fallback for unknown/default ids); other modes keep the existing behavior.
• 6 regression tests, including an end-to-end on_message test proving a >80-token output context survives on a Claude-only setup.

Testing

Full Python suite passes. Manually verified in JupyterLab on a Claude-only setup: attaching a 34-line cell output (well over the old 80-token budget) via "Ask about this output" and asking the agent to echo a sentinel marker + a derived value returned the correct SENTINEL-7F3A / col_sum_A=1198 — i.e. the full output reached the model. On main the same attachment comes back NO CONTEXT.

[pjdoland]

LGTM. Good catch that chat_model reflects the last non-Claude provider (or is None) in Claude Code mode, so the budget silently fell through to the 80-token floor and dropped output attachments. Deriving the limit from the configured Claude model via model_info_from_id is the right fix, and the end-to-end on_message test plus the sentinel round-trip verification make it convincing.