ENH: model unbonded solid-motor grain CM shift (#340) - #1138

I don't believe the solution could be that simple... But we can give it a try

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 84.34%. Comparing base (e0ff281) to head (ded59fe).
⚠️ Report is 72 commits behind head on develop.

@@             Coverage Diff             @@
##           develop    #1138      +/-   ##
===========================================
+ Coverage    82.18%   84.34%   +2.15%     
===========================================
  Files          122      130       +8     
  Lines        16355    17268     +913     
===========================================
+ Hits         13441    14564    +1123     
+ Misses        2914     2704     -210     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

Following up on my "I don't believe the solution could be that simple" comment — I worked through it and it does hold up. Recording the check so it is on the record:

The CM expression is exactly derivable. With grains packed against the fixed aft face and instantaneous pitch p = h + s, grain k's centre sits at z_k = z_aft + h/2 + k*(h + s). Averaging over k = 0..N-1:

mean(z_k)          = z_aft + h/2  + (h  + s)*(N-1)/2
mean(z_k) bonded   = z_aft + h0/2 + (h0 + s)*(N-1)/2
difference         = (h - h0)/2 + (h - h0)*(N-1)/2 = -(N/2) * (h0 - h)

which is precisely the implemented grains_center_of_mass_position - _csys * (grain_number/2) * (grain_initial_height - grain_height). The instantaneous pitch in the parallel-axis term is the right companion to that, and propellant_I_22 returns self.propellant_I_11, so the transverse components stay consistent for free.

The default path is unchanged. For grains_bonded=True, _grain_pitch_squared_sum() returns (h0 + s)**2 * sum(offsets**2), which is algebraically the same as the old np.sum(d**2) with d = linspace(-iv, iv, n) * (h0 + s). So existing motors get bit-identical numbers.

The remaining caveats are modelling choices rather than errors, and the docstring already states them: grain_separation is retained while the stack shortens, and there is no acceleration-dependent settling — aft packing is what happens under boost, and after burnout grain_height stops regressing so the CM freezes anyway, which lands on the right answer for the wrong reason. Fine as a documented first-order model.