[sarlinpe]

Summary

Caspar factor kernels generate outputs in distinct mathematical blocks separated by shared-memory reductions (FlushSumShared) and warp synchronizations:

1. Residuals (out_res, out_rTr)
2. Pose block (out_pose_jac, out_pose_njtr, out_pose_precond_*)
3. Calibration block (out_calib_jac, out_calib_njtr, out_calib_precond_*)
4. Point block (out_point_jac, out_point_njtr, out_point_precond_*)

Currently, a single flat expr_map is shared across all outputs. Common subexpressions are reused globally without barrier awareness. For example, in a SimpleRadial factor, calib_factor = -1.0 * (1 + k*r^2) was first computed during out_pose_jac into r54 and reused 350 lines later during out_calib_njtr across 12 FlushSumShared calls.

Under __launch_bounds__(1024, 1) in double precision, this 350-line cross-barrier live range triggered massive register spilling and stack slot clobbering in NVCC, causing calibration derivatives to diverge in COLMAP (see colmap/colmap#4799).

This PR scopes expr_map per output parameter group:

• base_expr_map stores read input symbols (camera pose, calibration parameters, 3D points, 2D pixels).
• When transitioning across output parameter blocks (e.g. pose -> calib -> point), expr_map is reset to base_expr_map.

This preserves full CSE within each parameter block while preventing intermediate derivative expressions from leaking across reduction barriers into subsequent parameter blocks.

Test Plan

• Passes existing Caspar factor codegen and unit tests.
• Verified in COLMAP's Caspar BA integration (Regenerate Caspar kernels with upstream SymForce fixes colmap/colmap#4799).

[sarlinpe]

Closing this PR: string-matching output name suffixes in dabseg_from_accessors.py is overly specific to factor output naming conventions, and furthermore do_cpse in the DAG optimizer undoes the scoping by deduplicating identical expressions globally across the entire DAG. A proper upstream solution should make the DAG optimizer / register scheduler barrier-aware.