OCR Changes Merged To `development` (2026-04-01 to 2026-04-03)¶

This note summarizes the OCR-facing changes already merged into development, centered on commit 489698e (deepseek reliability hardening).

Use it as a short operator/developer changelog for the April 1-3 rollout.

Runtime reliability¶

DeepSeek multi-GPU OCR now runs through a durable SQLite work queue instead of fragile fixed subprocess assignment.
Work items heartbeat while running and are requeued if a worker dies or goes stale.
Failed work items now default to one retry (max_attempts=2 total attempts), then become terminal failures for operator follow-up instead of bouncing forever.
Repair work is durable too: first-pass batches populate a second repair queue that workers drain after the main queue is empty.
Workers are launched in their own process groups so respawn can clean up orphaned runtime processes and recover GPU memory.

vLLM OCR now renders pages into memory and feeds a bounded render queue directly into inference, removing the temporary image-file round trip.
Rendering and inference overlap during the first pass.
Empty pages are detected before inference and skipped early.
Per-worker runtime JSON, GPU preflight output, GPU telemetry, durable queue state, and the final runtime summary now live under sidecars/ocr_runtime/.
Runtime summaries now expose steady-state inference timestamps so long-run throughput can be measured without startup noise.

Canonical OCR outputs remain one markdown/<stem>.md and one json/metrics/<stem>.metrics.json per source PDF.
Page boundaries are annotated with  comments alongside the page split markers.
Internal shard markdown and shard metrics move under sidecars/ocr_shards/ so downstream stages do not mistake them for canonical outputs.
If a repair retry hits the garbage cutoff again, GlossAPI now blanks that page slot instead of preserving the failed garbage text.
Repair queue durability and repair execution packing are separate concerns: queue accounting stays item-granular, while workers are allowed to combine multiple repair items into one larger execution batch.

The runner now auto-discovers wheel-managed CUDA libraries inside the selected DeepSeek virtualenv and prepends them to LD_LIBRARY_PATH.
Fresh A100 nodes should be validated first with:

python -m glossapi.scripts.deepseek_runtime_report --repo-root <repo-root>

This note only summarizes OCR work already merged into development.

It does not describe the still-in-progress branch work for: