Harden remote captioning and document GLM-5.3-Flash

[bghira]

Summary

• add a tested GLM-5.3-Flash vLLM deployment recipe and safer vLLM argument compatibility
• add optional image downscaling for OpenAI-compatible workers to avoid multimodal context overflows
• use trainingsample for Rust-backed image decoding and resizing, including fused API request batches
• allow WebDataset workers to defer decoding and avoid redundant per-sample image work
• support policy-rejection fallback prompts, endpoint timing visibility, and more robust retry handling
• harden WebDataset range reads and preserve partial progress when a work unit fails

Validation

• 492 tests passed
• 92% local patch coverage against main
• Python 3.12, Python 3.13, lint, and Codecov checks passed
• 128-image H100 rerun reduced measured data preparation from 48.2s to 36.8s (23.7%)
• reviewed diff contains no machine-specific paths or usernames

[codecov]

Codecov Report

❌ Patch coverage is 92.69406% with 16 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/caption_flow/workers/openai_compatible.py	88.98%	13 Missing ⚠️
src/caption_flow/processors/webdataset.py	96.87%	1 Missing ⚠️
src/caption_flow/utils/chunk_tracker.py	94.44%	1 Missing ⚠️
src/caption_flow/utils/image_processor.py	97.91%	1 Missing ⚠️

Files with missing lines	Coverage Δ
src/caption_flow/models.py	96.13% <100.00%> (+0.04%)	⬆️
src/caption_flow/utils/vllm_config.py	93.33% <ø> (ø)
src/caption_flow/workers/caption.py	63.52% <100.00%> (ø)
src/caption_flow/processors/webdataset.py	90.44% <96.87%> (+0.42%)	⬆️
src/caption_flow/utils/chunk_tracker.py	77.91% <94.44%> (+0.77%)	⬆️
src/caption_flow/utils/image_processor.py	90.54% <97.91%> (+16.46%)	⬆️
src/caption_flow/workers/openai_compatible.py	90.52% <88.98%> (+0.22%)	⬆️

... and 3 files with indirect coverage changes

🚀 New features to boost your workflow:

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