Add pinhole camera projection to reorder texture

[csparker247]

Summary

Adds a pinhole camera projection model to ReorderUnorganizedTexture, alongside the existing orthographic ("snapshot from above") sampling. Instead of casting parallel rays along the mesh's shortest axis, ProjectionMode::Camera casts perspective rays from a camera center, rendering the textured mesh as if photographed through a pinhole camera.

The camera is either:

• supplied explicitly via a --camera-file (intrinsics fx/fy/cx/cy, output width/height, and a world-to-camera pose), or
• auto-derived to frame the mesh — positioned along the shortest OBB axis, looking at the centroid, with intrinsics/output size chosen to preserve the input texture's pixel density.

Pipeline integration

Camera mode runs through the standard render graph rather than a separate code path. ReorderTextureNode gains two input ports:

• projectionMode → setProjectionMode()
• projectionParams → setProjectionParams()

Both are serialized into the node cache (NLOHMANN_JSON_SERIALIZE_ENUM for the mode; to_json/from_json for the params struct, with the 4x4 pose flattened to 16 doubles). The rt_reorder_texture app parses the new options and wires them onto the node; the camera-file parsing lives in a dedicated static ParseCameraFile() helper that returns std::optional<ProjectionParams>.

Output handling

The output node branches on the requested file extension (works for both projection modes):

• OBJ → writes the textured mesh (MeshWriteNode)
• image (jpg/jpeg/png/tif/tiff) → writes just the reordered texture (WriteImageNode)

The -o option is renamed --output-mesh → --output-file to reflect that it now accepts either.

Additional outputs

• New per-pixel 3D position map (CV_32FC3) output / --position-map option and node port. Camera-mode positions are in the mesh's world frame; orthographic positions are in the realigned sampling frame. Pixels with no intersection are NaN.
• Depth map semantics documented for both modes (orthographic: distance from the sampling plane; camera: optical-axis depth / camera-space Z).

Other

• TextureDewarp: switch the dewarp sampling to OutputWidth at 8192 px for higher-resolution output.

Notes

• Camera mode shares all outputs (texture, UV map, depth, position) with orthographic mode, so it flows through the same graph; the output writer is selected by extension as described above.
• No textured-OBJ fixture exists in the repo, so this was verified by full build + CLI help; the ray/sampling math in create_texture_camera_() is exercised through the node rather than an end-to-end render test.

🤖 Generated with Claude Code

[csparker247]

Addressed the review findings in four focused commits:

• bb628cb Decouple setProjectionParams from mode — it no longer force-sets Camera; added clearProjectionParams() to revert to auto-derivation. Node deserialize_ now restores params before mode (serialized mode is authoritative) and clears stale params when none were cached. (Fixes the serialize/deserialize coupling.)
• c07cf2d Validate pinhole camera params — new rt::ValidateProjectionParams() rejects non-positive/non-finite fx/fy, non-positive image size, non-finite principal point, and extrinsics whose rotation block isn't a proper rotation (R·Rᵀ ≈ I and det ≈ +1). Wired into create_texture_camera_() (covers the degenerate auto-camera too) and ParseCameraFile() (clean CLI error). Added a GoogleTest (rt_TestReorderUnorganizedTexture, 7 cases). (Fixes fx/fy guard + non-orthonormal pose + zero-extent auto-camera.)
• 33db713 Share color sampling — extracted the duplicated per-hit UV→color block into sample_surface_color_(); each loop hoists one inputTexture_.empty() check so neither path calls getRectSubPix on an empty image (depth/position still produced). Documented the no-near-plane-clipping limitation in CreateProjectiveUVMap. (Fixes dedup + empty-texture.)
• 60d60b0 CLI polish — warn when --camera-file is set without --projection camera (+ note in help); note float-capable format (.tif) in --depth-map/--position-map help.

On the float-map finding: WriteImage already warns when writing float to a non-float format and routes .tif through a float-capable writer, so no code-path change was needed beyond the help note.

Full build is clean and all 4 tests pass.