Timing observations: do_code_enrichment + generate_picture_images on a 689-page PDF

### Summary

Timing data from converting a 689-page PDF under different
`do_code_enrichment` / `generate_picture_images` configurations. Sharing in
case it's useful for triage.

This is a single-document observation from an internal evaluation pilot, not
a systematic benchmark.

### Environment

- docling 2.119.0
- Python 3.12.14
- macOS, 24GB RAM
- Test document: 689-page PDF, 47.6MB (SAP training material, born-digital
with embedded ABAP code blocks, tables, and figures)

### Measurements

| Config | Scope | Time | Notes |
|---|---|---|---|
| Baseline (no image export, no code enrichment) | 689 pages (full doc) | 654s (~10.9 min) | Peak RSS 6.68GB, peak memory footprint 9.9GB, exit code 0 |
| `do_code_enrichment=True` only | 1 page with code | 86s | vs 18s for the same page without it |
| `generate_picture_images=True` only | 1 page with a figure | 23s | No `do_code_enrichment` |
| `do_code_enrichment=True` + `generate_picture_images=True` together | 36 pages | >18 min, killed before completion | Not run to completion; the full 689-page document was not attempted under this configuration |

### Question

Is this a known interaction?

Happy to share the test script / more detailed timing breakdown if useful.

Begin by reproducing the four reported configurations with the 689-page, 47.6MB PDF under Python 3.12.14, comparing elapsed time and peak memory. Focus on the interaction between do_code_enrichment and generate_picture_images, and use the single-page and 36-page measurements as checkpoints. Done means establishing whether the combined slowdown is reproducible and providing actionable triage findings.

Written by the indexing model from the issue text.