feat(experiment-05): VLM + OCR hybrid provider (gpt-5.4) + canonical 40-sample run

Two-leg hybrid extraction: Azure DI prebuilt-layout produces markdown +
bbox layout, then Azure OpenAI gpt-5.4 reads the document image alongside
the OCR markdown with a strict-mode JSON Schema response_format and a
trust-hierarchy system prompt ("prefer the image when image and OCR text
disagree"). Mapper folds the upstream layout's word/line polygons into
OCRResult, closing the bbox gap E04 documented.

Per the user's session-start scope reduction, only variant 1 + gpt-5.4
is implemented; variant 2/3 + gpt-4o/gpt-5 deferred (the plumbing is in
place — variant 3 is `params.includeBboxAnnotations: true`, deployment
swaps via `params.azureOpenAiDeployment`).

Canonical run (cb677a90, 40 samples): pass_rate 0.975 (39/40),
f1.median 0.965 (tied with CU), f1.mean 0.941 (best across all 5
experiments), recall.mean 0.917 (best). Hybrid is the production-stack
winner on this dataset.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: strukalex

apps/backend-services/src/workflow/activity-registry.ts

@@ -47,6 +47,14 @@ export const REGISTERED_ACTIVITY_TYPES: Record<string, RegisteredActivityType> =
       description:
         "VLM-direct extraction (Azure OpenAI chat completions with vision input + strict JSON schema response_format)",
     },
+    "vlmOcrHybrid.extract": {
+      description:
+        "VLM + OCR hybrid extraction (Azure DI prebuilt-layout markdown + Azure OpenAI chat completions with vision + strict JSON schema response_format)",
+    },
+    "azureOcr.readPlain": {
+      description:
+        "Azure DI prebuilt-layout (markdown content + per-line/per-word polygons; no field extraction). Sync submit+poll wrapper for the VLM hybrid pre-pass.",
+    },
     "ocr.cleanup": { description: "Post-OCR text normalization" },
     "ocr.enrich": {
       description: "Enrich OCR results using field schema and optional LLM",