Add NemotronParseBBox.to_original_coordinates for raw nemotron-parse output

[jamesbraza]

What

nemotron-parse emits bounding boxes normalized to its letterboxed 2048×1648 canvas (aspect-preserving resize + centered white padding). NVIDIA's hosted NIM maps these back to the input image server-side — via postprocessing.py::transform_bbox_to_original — before returning, so PaperQA's existing decode (coord × input_dims) is correct against NIM. Raw vLLM / Hugging Face output is not transformed, so a consumer of raw model output draws every box shifted and compressed — worst at low DPI, where the canvas padding is largest.

This ports that inverse transform so raw-output consumers can recover original-image coordinates:

• NEMOTRON_PARSE_TARGET_HEIGHT = 2048, NEMOTRON_PARSE_TARGET_WIDTH = 1648 — sourced (with a link) from the model's preprocessor_config.json
• transform_canvas_bbox_to_original(...) module function
• NemotronParseBBox.to_original_coordinates(height, width) method, mirroring the existing to_page_coordinates

Why it lives here

Our self-hosted vLLM Modal proxy returns raw canvas coordinates rather than replicating NIM's server-side mapping; this is the canonical, tested transform it will call. PaperQA's NIM decode path is intentionally unchanged — applying this on NIM coordinates would double-transform.

Tests

test_api.py adds:

• Parity against NVIDIA's reference transform_bbox_to_original across the no-resize, downsized, and width-resized regimes
• Round-trip recovery (forward-letterbox a page rectangle → invert → recover, 0 px error)
• Low-DPI heavy-padding case, where the correction is large vs. naive page scaling

🤖 Generated with Claude Code

[Copilot]

Pull request overview

Adds a faithful inverse “letterbox + pad” transform for raw nemotron-parse outputs whose bounding boxes are normalized to the model’s fixed 2048×1648 canvas, enabling callers (e.g., self-hosted vLLM/HF deployments) to recover original-image pixel coordinates without affecting the existing NVIDIA NIM decode path.

Changes:

• Introduces NEMOTRON_PARSE_TARGET_HEIGHT/WIDTH and transform_canvas_bbox_to_original(...) to invert nemotron-parse preprocessing.
• Adds NemotronParseBBox.to_original_coordinates(...) as a companion to to_page_coordinates(...).
• Extends test_api.py with parity tests vs. NVIDIA’s reference implementation plus round-trip and low-DPI padding regression coverage.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
packages/paper-qa-nemotron/src/paperqa_nemotron/api.py	Adds canvas-dimension constants, the inverse-transform helper, and a NemotronParseBBox method to map canvas-normalized bboxes back to original-image pixels.
packages/paper-qa-nemotron/tests/test_api.py	Adds reference-parity, round-trip, and heavy-padding regression tests for to_original_coordinates.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[jamesbraza]

Claude why