(WIP) Add per-link DEM error and velocity estimation

[scottstanie]

(Opening to make others aware, but some cleanup should probably be done)

I've been playing with an addition to spurt that @piyushrpt described a long time ago, but we never added her: During the temporal unwrapping stage, a per-link phase velocity and dem-error can be estimated and removed before spatially unwrapping. The makes a noticeable difference on some high-res urban scenes with large sections of layover and big jumps due to DEM errors (e.g. the edge of the top of a sky scraper will be coherence, but the phase is totally unrelated to the next pixel over on the ground).

This has a lot in it, but the main things are

• Design matrix now accommodates the velocity + the DEM error term
• The workflow accepts a --baseline-csv file listing the perpendicular baseline terms. this is the most open to reworking, since i just sort of picked a format that made sense and matched my own current processor.
• There are new grid search parameters to test a range of velocities and $\Delta h$ terms

The fullest way to test would be to use my dolphin branch here, which added the plumbing to dolphin's unwrapper options.

ok now the claude summary:

---

What's new

Feature:

• io/_baseline.py — CSV loader for perpendicular baselines. Accepts either per-SLC (date,bperp_m) or per-IFG (reference,secondary,...,bperp_m) format. Per-IFG is auto-detected and converted via least-squares with
  SLC[0] as reference.
• links/_design_matrix.py — builds the (nifgs, 2) design matrix from wavelength, slant range, look angle, and perpendicular baselines. Column 0: velocity sensitivity (rad per mm/yr). Column 1: DEM-error
  sensitivity (rad per m). Uses sin(θ) for the DEM-error term (previously reviewed & corrected on-branch).
• links/_grid_search.py — GridSearchLinearModel with vectorized estimation: precomputes E = exp(1j · design @ grid) once, batches all links into a single BLAS E.T @ V matmul, then replaces per-link Nelder-Mead
  with a quadratic fit over the 3×3 coherence stencil around the best grid point. Fully vectorized across links.
• workflows/emcf/_solver.py — new link_model plumbing, link_params/link_coherence outputs, and integrate_link_params() (sparse-incidence LSQR) for converting per-link gradients to per-point values.
• workflows/emcf/_output.py — writes velocity_mm_yr.tif, dem_error_m.tif, link_model_coherence.tif aligned to the input raster grid.

New CLI args (all additive, all default to current behavior):

• --baseline-csv — path to CSV; enables the feature
• --no-velocity-estimation — hard opt-out
• --wavelength, --slant-range, --look-angle-deg — sensor geometry
• --velocity-range MIN MAX STEP, --dem-error-range MIN MAX STEP — grid

New settings type: LinkModelSettings with validation (wavelength > 0, look angle ∈ (0, 90), etc.).

Docs:

• docs/theory/dem-error-velocity.md — derivation of the design matrix
• docs/notebooks/dem_error_estimation.ipynb — end-to-end tutorial

Tests added (~1100 lines):

• test/io/test_baseline.py — both CSV formats, error paths
• test/links/test_design_matrix.py — shape + sensitivity values
• test/links/test_grid_search.py — vectorized-search correctness
• test/workflows/test_emcf_baseline_integration.py — end-to-end CLI
• test/workflows/test_emcf_link_model.py — solver integration + integrate_link_params

Bundled fixes (called out so reviewers know they aren't DEM-error-specific)

• macOS multiprocessing memory explosion (a8b29fa): worker pools in mcf/_ortools.py and workflows/emcf/_solver.py switched from default fork to forkserver + initializer, so constant arrays (dual graph, costs,
  solver state) are shared once per worker instead of inheriting the parent's entire address space. On macOS, COW sharing breaks down under Python refcounting, causing per-worker RSS to balloon. This is a
  general-purpose fix and would be valuable independent of this PR.
• Baseline CSV loading flexibility (69efdf9): tolerant of column order + extra columns.
• Logging clarifications (3bc8211): a couple of logger.info messages made less ambiguous during tile unwrap.

How to use

python -m spurt.workflows.emcf
-i /path/to/phase-linked-stack
--baseline-csv /path/to/baselines.csv
--wavelength 0.031
--look-angle-deg 35

Outputs appear in the output folder:

• velocity_mm_yr.tif
• dem_error_m.tif
• link_model_coherence.tif
• standard unwrapped interferograms (unchanged)

Things reviewers should look at

1. links/_grid_search.py — this is where the per-link estimation happens, and where the perf work lives. The quadratic-refinement replacement of Nelder-Mead is the biggest behavioral change worth a careful read.
2. Sign convention in links/_design_matrix.py — sin(θ) vs cos(θ) for DEM-error sensitivity was flipped on-branch (42486ad); the current form follows the convention in the theory doc.
3. workflows/emcf/_solver.py integration point — the model is applied via phase_diff(..., model=model_pred) which returns the full (not residual) gradient wrapped around the model. Confirmed in comments at
   _solver.py:178-180.
4. forkserver switch — behavior is identical on Linux; only the startup path differs. Worth sanity-checking on a Linux CI run.