fix(mind): support json-object-only providers

[keli-wen]

Summary

• Support json_object-only providers (DeepSeek, ...) for both structured-output stages — draft structuring (flows/paper/_structure.py) and agentic retrieval (mind/retrieval.py) — without degrading capable providers or adding a provider registry.
• Add quantmind.utils.structured_output.run_structured: a strict-first fallback ladder shared by both call sites. It runs the SDK's native strict json_schema structured output and, only when a provider rejects that response_format at request time (a narrow openai.BadRequestError check), re-runs the same call once in JSON-object mode with local Pydantic validation. Callers pass only cfg.model; the fallback is invisible, and bad requests unrelated to response_format are re-raised, never swallowed.
• The helper is layer-agnostic: each call site supplies its own build_agent (how the agent is constructed) and run (which runner executes it) closures, so flows keeps run_with_observability and mind keeps Runner.run — no mind -> flows import edge is introduced.

Root cause

output_type= always makes the Agents SDK send a strict response_format={"type": "json_schema"}. Native OpenAI routes accept it; some LiteLLM-routed providers (DeepSeek) reject that response_format outright at request time, before any output exists — so a parse-stage retry can't help; the request itself must change.

What changed since the first version of this PR

The first iteration detected incapability by a model.startswith("litellm/") prefix and forced every routed model into json_object mode. That downgraded capable providers (losing schema enforcement) and, because json_object gives no schema guarantee, required a growing salvage stack: a *_node_ids alias normalizer, a UUID regex scan over free text, and a second tool-less selector agent. This revision detects incapability by the provider's own rejection instead — restoring strict schema for every capable provider (including the GPT baseline, which now genuinely exercises the strict path) and deleting that salvage stack. Net -92 lines in mind/retrieval.py.

Trade-off

A json_object-only provider now pays one rejected first request per call (the strict attempt). No provider registry or capability table is introduced (see the design doc's Out-of-Scope). A per-process "this model rejected strict" memo could remove that cost later if it matters.

Validation

• bash scripts/verify.sh — 394 passed, 85.34% coverage.
• Offline (both stages): strict-is-the-default, strict→json_object fallback on a simulated response_format rejection, and an unrelated BadRequestError propagating unswallowed.
• Live: scripts/verify_structure_e2e.py with OPENROUTER_API_KEY — OpenRouter DeepSeek V4 Flash exercises the JSON-object fallback, GPT-5.6 Luna is the strict-path baseline.

Closes #126