feat: Logprobs API + structural tag constraint for tool calls

[raullenchai]

Summary

• Logprobs API (fix: init crash, JSON corruption, GC leak, cloud routing gaps #11): Propagate mlx-lm per-token logprobs through the full stack (StreamingOutput → GenerationOutput → API response). Supports both streaming and non-streaming chat/completion endpoints with logprobs: true and top_logprobs: 0-20 per OpenAI spec.
• Structural tag constraint (feat: TTFT cache fix, MiniMax reasoning parser, logprobs API, tool logits #7): Extend MiniMaxToolLogitsProcessor with parameter value schema tracking — biases toward valid JSON types (string/number/boolean/object/array) at the start of parameter values during generation. SimpleEngine gets post-generation validation with warning logs for schema mismatches.
• Fix: Prevent server crash from malformed response_format schemas (wraps build_json_system_prompt and parse_json_output in try/except).

Changes

File	Change
api/models.py	Add TopLogProb, TokenLogProb, ChoiceLogProbs models; add logprobs/top_logprobs to request/response types
models/llm.py	Pass response.logprobs (mx.array) through StreamingOutput
engine/base.py	Add logprobs: Any to GenerationOutput
engine/simple.py	Propagate logprobs + token ID in streaming
api/tool_logits.py	Parameter value state tracking, JSON type bias, validate_param_value(), _extract_param_schemas()
server.py	_extract_token_logprob() helper, top_logprobs validation, streaming/non-streaming logprobs wiring, _validate_tool_call_params() post-generation check
api/tool_calling.py	Wrap json.dumps(schema) to prevent crash on non-serializable schemas

Test plan

• pytest tests/test_api_models.py — model schema tests pass (72 tests)
• pytest tests/test_tool_logits.py — tool logits processor tests pass (9 tests)
• pytest tests/test_server.py — server endpoint tests pass (34 tests)
• Full test suite: 904 passed, 14 skipped (4 pre-existing failures: missing torch/async framework)
• Manual: curl with logprobs: true, top_logprobs: 5 returns token-level logprobs
• Manual: streaming with logprobs returns per-chunk logprobs
• Manual: tool calls with schema-typed params get correct JSON bias

🤖 Generated with Claude Code