Share caption policy and retries across native vLLM and API workers

[bghira]

Summary

• Move caption acceptance, JSON validation/repair, refusal matching, semantic retries, and stage progression into a backend-independent pipeline used by both workers.
• Add shared inference/stage output_processing, response_format, and retry_response_format configuration. Keep existing API-worker policy flags as compatibility aliases.
• Support native vLLM constrained decoding through both structured-output and older guided-decoding APIs, preserving native sampling controls and isolated retry overrides.
• Fix native hot reload retaining stale stage settings, unsuccessful items reaching later stages, and decoded images being discarded before generation. Preserve original dimensions through resize recovery.
• Reuse encoded API images across stages and release request buffers after each batch. Update the recipe and example configuration.

Validation

• Full suite: 541 passed, 2 skipped.
• Shared stage runner: 100% statement coverage; locally calculated executable-line patch coverage approximately 99%.
• Focused Ruff checks and git diff --check pass.
• Cross-backend regression tests cover malformed JSON, semantic retries, constrained decoding, native sampling, stage failures, config reload, and image-buffer reuse.
• Native vLLM calls are mocked in the CPU test suite; no GPU throughput comparison is claimed.

No deployments, worker restarts, or changes to running caption jobs.

[codecov]

Codecov Report

❌ Patch coverage is 98.91892% with 4 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/caption_flow/workers/caption.py	97.97%	2 Missing ⚠️
src/caption_flow/utils/output_policy.py	99.27%	1 Missing ⚠️
src/caption_flow/utils/vllm_config.py	94.44%	1 Missing ⚠️

Files with missing lines	Coverage Δ
src/caption_flow/models.py	96.21% <100.00%> (+0.06%)	⬆️
src/caption_flow/utils/image_processor.py	98.66% <100.00%> (+8.12%)	⬆️
src/caption_flow/workers/openai_compatible.py	91.75% <100.00%> (-0.62%)	⬇️
src/caption_flow/workers/pipeline.py	100.00% <100.00%> (ø)
src/caption_flow/utils/output_policy.py	99.27% <99.27%> (ø)
src/caption_flow/utils/vllm_config.py	96.66% <94.44%> (+3.33%)	⬆️
src/caption_flow/workers/caption.py	74.55% <97.97%> (+10.70%)	⬆️

... and 1 file with indirect coverage changes

🚀 New features to boost your workflow:

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