fix(local-vlm): apply json schema guided decoding

[rikkurea-boop]

Summary

Apply caller-provided JSON schemas to the Local VLM single-image generation path so structured-output requests reach the actual vLLM runtime configuration.

Problem

LocalVLMServing_vllm.generate_from_input() accepts json_schema, but the value is not forwarded to the single-image path.

As a result:

generate_from_input(json_schema=...)
→ generate_from_input_one_image()
→ llm.generate(...)

still uses the default self.sampling_params, so the requested guided-decoding schema is lost before runtime generation.

Fix

Forward json_schema into the single-image path.

When a schema is provided:

• create call-local GuidedDecodingParams(json=json_schema);
• create call-local SamplingParams;
• preserve the existing temperature, top-p, max tokens, top-k, repetition penalty, and seed settings.

When no schema is supplied, continue reusing the original self.sampling_params object.

Tests

The regression exercises the real public:

LocalVLMServing_vllm.generate_from_input()

through the real single-image implementation and observes the sampling parameters passed to llm.generate().

Pre-fix, the engine receives the default sampling object even when a schema is supplied.

Post-fix:

sampling_params.guided_decoding.json == caller_schema

The regression also verifies:

• existing sampling knobs are preserved;
• the no-schema path continues to reuse the default sampling object.

Heavy vLLM runtime components are stubbed only below the defect boundary with explicit compatible interfaces.

Validation

• targeted RED: 1 failed, 1 passed;
• targeted GREEN: 2 passed;
• full CPU/API suite passed on Ubuntu Python 3.10/3.11/3.12/3.13;
• full CPU/API suite passed on Windows Python 3.10/3.12;
• changed-file py_compile: passed;
• git diff --check: passed.

Scope

Only these paths are changed:

• dataflow/serving/local_model_vlm_serving.py
• test/cpu_only/test_local_vlm_json_schema_runtime.py

No P09.07, P11, G6, P09.05, or unrelated serving changes are included.