kapoorkarmanya/cv-accuracy-upgrades

feat: add ECC alignment and Hough circle snapping to improve accuracy

removed the wrong changes made by the incorrect prompts

Fix the Supabase empty array timeout bug and the Integer/String type mismatch causing duplicate sheet insertions.

Unify the duplicate detection logic so both Barcode (sheet_id) and Roll Number (omr_id) collisions are stashed for user resolution, while bypassing Identity Errors.

feat(scanner): implement non-blocking batch conflict resolution

Locate and close the orphaned `try` block in `main.py` that is causing the SyntaxError around the `/generate-omr` endpoint.

refactor(omr-generation): harden PDF generation route, proxy handler, and backend memory buffers

feat(omr-engine): implement statistical dynamic calibration and out-of-scope boundary filters

feat(taxonomy): implement dynamic admin-controlled test categorization" -m "Key updates:

feat(billing): implement custom tier escrow engine and cashfree webhook security" -m "Architecture updates:

Implement Enterprise Quota UX in the Scanner Client. Prevent users from submitting batches that exceed their limit, and handle the 402 Error gracefully.

Remove the slow `PyPDF2` loop inside `/download-omr` and wire the endpoint directly to the native, pixel-perfect `generate_omr.py` script.

Wire the `/download-omr` endpoint to use the existing, pixel-perfect `generate_omr.py` script instead of the slow PyPDF2 fallback.

Implement the Pipelined Assembly Line to overlap upload and grading, and add a Redis connection safeguard for the training sync task.

Completely replace `sync_training_data_task` to upload a single zip archive instead of individual PNGs, bypassing Supabase rate limits.

Completely replace `sync_training_data_task` to create a single zip archive of patches instead of uploading individual PNGs, preventing Supabase DDOS errors.

fix(engine): resolve redis collision and supabase rate limits

perf: zip training data to prevent rate limits

perf: offload audit image uploads to background queue for zero-latency VIP grading

perf: completely drop empty and metadata bubbles from training pipeline to maximize VIP speed

moving the 2% empty bubble rate limiter upstream into ision.py before local disk writes

perf: move result saving to VIP queue to prevent upload blocking

fix: forcefully remove redis lock and rewrite training patch upload

fix: resolve redis lock ssl syntax and implement local disk handoff

fix: catch both strict and implicit celery task names in routing

fix: resolve upstash celery ssl requirements and add health check

fix: add huggingface configuration readme

Production deployment: Cleaned Binaries