fix(llm): recover from empty responses caused by reasoning token budgets

[ajianaz]

Fixes #536

What

• ChatResponse now captures finish_reason and reasoning_content (string or parts array) from LLM responses
• Empty content + finish_reason=length triggers an automatic retry with doubled max_tokens (…→32768 ceiling) inside chat_completion
• Last-resort salvage: JSON-looking reasoning text is used as the raw response (parse layer still validates it)
• Parse layer reports an explicit "provider returned an EMPTY response" error instead of serde EOF noise
• Default max_tokens raised 4096 → 8192

Why

Reasoning models (e.g. glm/glm-5.3 via Bifrost) can spend the entire output budget on chain-of-thought and return HTTP 200 with content: "" and finish_reason: "length". cora never read those fields, silently fed an empty string to the JSON parser, and surfaced the misleading EOF while parsing a value at line 1 column 0 — while the model's actual findings JSON sat in the backend logs (codecoradev/cosy#63). The old retry re-sent the same 4096 budget, guaranteeing identical failure.

Testing

• New unit tests: budget escalation table (length/stop/none × 32768 ceiling), reasoning salvage (plain JSON / fenced / parts array / prose / none), explicit empty-raw parse error
• Full suite green (912 tests, tree-sitter feature); cargo fmt clean; no new clippy findings vs develop
• Pre-commit hook (Cora review) passed locally

All commits signed-off (DCO).