fix: projection convergence loop referenced previous time step's energy

[MZKC]

Summary

The projection routine re-converges the projection orbitals (Houston /
instantaneous eigenstates) at each output step with a CG loop, exiting when the
kinetic-energy change dE = energy%E_kin - rt%E_old drops below
threshold_projection. rt%E_old is a persistent (s_rt) value carried over
from the previous time step, and it was not reset to the current step's starting
E_kin before the loop. The first iteration therefore measured dE against the
previous time step rather than against the current step's starting orbitals.

When E_kin after a single CG step happens to land within threshold_projection
of the previous step's value, the loop can exit after one iteration without having
converged the projection orbitals — a premature exit. This is a coincidence, so it
shows up intermittently and is more likely for large systems / long runs (matching
the report that the loop "sometimes" stops without converging).

Fix

Reset rt%E_old to the current step's E_kin before the loop, and update it
before the exit test so it holds the converged value for the next step's
"already converged" cross-step check. dE inside the loop is now the change
within this projection's own iterations.

Verification (Wisteria-O)

GS → RT with projection_option='gs', comparing the fixed binary against the
pre-fix (develop-2.0.0) binary:

• cubic Si (8 atoms, 2x2x2 k) and SiO2 (18 atoms, 4x4x4 k,
  I = 1e14 W/cm^2):
  • RT dynamics are unchanged. Si_rt_energy.data / SiO2_rt_energy.data
    are identical (Eall bit-for-bit); the macroscopic current differs only at
    the 1e-19–1e-21 (denormal) level. The projection is a read-only analysis,
    and the fix does not alter the dynamics.
  • The projection converged in the same number of iterations for both
    binaries; the per-iteration dE for iter >= 2 is bit-for-bit identical, and
    only the first iteration's printed dE changes (reflecting the corrected
    reference). The premature exit is a rare coincidence that did not trigger in
    these short runs, so base and fix converged identically.

The change is therefore a safe correctness fix: it removes a rare premature exit
(the reported intermittent under-convergence) and is a no-op in the common case.

🤖 Generated with Claude Code

[jenkins-diana]

Can one of the admins verify this patch?

[MZKC]

Reviewed — no correctness issues found. Seeding rt%E_old = energy%E_kin before the projection CG loop is the right fix: the first iteration's dE is now measured within this step (against the freshly computed starting-orbital E_kin) instead of against the previous time step, so the loop no longer exits after a single CG step. Moving the in-loop rt%E_old = energy%E_kin ahead of the exit check is behavior-neutral.