johanjoensson / impurityModel / 35281217541

94%

Build Type

push

github

Committed by johanjoensson

Commit Message

feat: report the RIXS intermediate-resolvent and final-state basis maxima

RIXS reaches two bases through three solver tiers, and in both cases the obvious
read is wrong. `KrylovShiftedResolvent.solve` leaves `tmp_basis` at the right-hand
side's support and tracks the recurrence only in its cap proxy; the tensor path's
out-resolvent cache serves a hit *without* regrowing `green_basis`, which was
cleared at the top of the unit, so reading it back gives 0 for a solve that really
ran on the cached sector. Both were measured, not reasoned about: instrumenting
them made the two RIXS drivers disagree on identical physics.

So nothing reads a basis object back. Each R1 tier reports its own support -- the
spectral cache its sector, the recycler its proxy count, the BiCGSTAB fallback the
basis it genuinely grew in place -- and `eval_out` returns the final-state support
it used beside its value. A unit with any unmeasured solve says "not tracked"
rather than offering a lower bound over an unknown subset of its wIn points.

The two bases are reported separately: they are reached by different solvers and
collapsing them into one number would invite comparing them. Rows are keyed by
eigenstate, the wIn chunks of one eigenstate collapsing into a row the same way
eigenstate chunks do elsewhere.

The accumulator is owned by `calc_map`/`calc_tensor_map` rather than the flat
driver, because the adaptive wIn sampler runs that driver once per refinement pass
and the report belongs to the finished map.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

Coverage Stats

139 of 141 new or added lines in 2 files covered. (98.58%)

98 existing lines in 4 files now uncovered.

34893 of 38058 relevant lines covered (91.68%)

0.92 hits per line