chore: add design_row assertion and update spec per code review

- Add assert(design_row == n_meas) in ComputeGeodesicAcceleration
- Remove v_acceleration_ and geodesic_alpha_ from spec (not used)
- Clarify that direction sets (D) are deferred, B/K azimuths implemented

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: daleroberts

docs/superpowers/specs/2026-03-11-geodesic-acceleration-design.md

@@ -15,7 +15,7 @@ Add geodesic acceleration (Transtrum & Sethna, 2012) to the existing LM solver i
 ## Design Decisions
 
 - **Analytic second derivatives** (not finite-difference) — exact, no step-size tuning, but requires per-measurement-type implementation.
-- **High-value types first** — distances (S, C, E, M) and directions/azimuths (D, B, K). GPS baselines are nearly linear in the ECEF Cartesian parameterisation; other types can be added later.
+- **High-value types first** — distances (S, C, E, M) and individual azimuths (B, K). Direction sets (D) are deferred: they use derived angles (differences of constituent directions) requiring more complex second-derivative logic. GPS baselines are nearly linear in the ECEF Cartesian parameterisation; other types can be added later.
 - **Separate acceleration pass** — new `ComputeGeodesicAcceleration()` function called after existing `SolveLM()`. Keeps `SolveLM()` untouched.
 - **Lambda decay unchanged** — geodesic acceleration is additive to the current simple decay strategy. Gain-ratio adaptation is a separate future enhancement.
 - **Opt-in via CLI flag** — `--lm-geodesic` (requires `--lm-enabled`). Default off.
@@ -174,12 +174,12 @@ void ComputeGeodesicAcceleration(const UINT32& block);
 
 **New member variables:**
 ```cpp
-vector<matrix_2d> v_acceleration_;
 vector<matrix_2d> v_choleskyFactor_;   // saved Cholesky factor for acceleration solve
 vector<vector<double>> v_scaleDiag_;   // saved scaling vector (when scale_normals_to_unity active)
-double geodesic_alpha_;
 ```
 
+Note: The acceleration vector is computed as a local variable within `ComputeGeodesicAcceleration()` rather than stored as a member — simpler with no persistent state needed. The alpha parameter is read from `projectSettings_.a.lm_geodesic_alpha`.
+
 ### Modified Iteration Loop
 
 After existing `SolveLM()` call:

dynadjust/dynadjust/dnaadjust/dnaadjust.cpp

@@ -7056,6 +7056,8 @@ void dna_adjust::ComputeGeodesicAcceleration(const UINT32& block)
 		}
 	}
 
+	assert(design_row == n_meas && "ComputeGeodesicAcceleration: design_row mismatch");
+
 	// Step 3: Assemble RHS: rhs = -(AtVinv * f_vv)
 	matrix_2d accel_rhs(n_params, 1);
 	AssembleAccelerationRHS(v_AtVinv_.at(block), f_vv, accel_rhs);