Validate and recover structured captions

[bghira]

Add structured JSON validation and semantic recovery to the OpenAI-compatible worker, based on the behavior exercised by the Ideogram-style captioning run.

• optionally reject malformed JSON, including non-standard NaN and Infinity
• conservatively repair invalid JSON escape sequences
• canonicalize JSON, correct inverted [ymin, xmin, ymax, xmax] boxes, and remove exact duplicate elements
• treat invalid structured output as retryable caption failure
• support retry-specific standard sampling through shared stage configuration
• support worker-local retry_extra_body for provider-specific recovery schemas
• validate structured-output and retry configuration eagerly
• document the complete primary and recovery setup without deployment-specific paths

The JSON validator checks syntax, while provider response_format remains responsible for schema-constrained decoding. Bounding-box normalization only corrects coordinate ordering; it does not infer or score boxes.

Validation:

• 499 tests passed
• 100% changed-line coverage (131/131 lines)
• Black formatting check passed
• targeted Ruff correctness checks passed
• all documented YAML examples parse successfully

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

Files with missing lines	Coverage Δ
src/caption_flow/models.py	96.15% <100.00%> (+0.02%)	⬆️
src/caption_flow/utils/vllm_config.py	93.33% <ø> (ø)
src/caption_flow/workers/caption.py	63.85% <100.00%> (+0.32%)	⬆️
src/caption_flow/workers/openai_compatible.py	92.37% <100.00%> (+1.50%)	⬆️

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.