FFS - EXPERIMENT - Kinetic harness attribution: branch is bit-neutral vs #398; deltas inherited

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Author: logan-nc

handoff/issue376/RESULTS.md

@@ -1145,6 +1145,13 @@ gold — again equilibrium-resolution-dominated at these sizes. The 1e-3 certifi
 extra resonance refinement on this case; the withheld-seed aliasing stress test has **not** been
 run (rational windows are always seeded by construction) and is recorded as outstanding.
 
+**Kinetic-case harness attribution**: vs develop the three `solovev_kinetic_*` cases move
+substantially (et[1] 1.7–2.5%, Im up to 23%, NTV torque magnitude, and NTV ψ-quadrature
+evaluations 840 → 60 — the torque quadrature was apparently spending 14× the evaluations
+resolving geometry noise). Vs the #398 branch: **zero changed quantities (34/34 unchanged)** —
+the deltas are inherited entirely from route (a)'s Solovev geometry fix, and this branch is
+bit-neutral on the kinetic cases with the knob off.
+
 Status: committed on `performance/decoupled-el-matrix-grid` (2225813fc), **default OFF**
 (`kinetic_grid_tol = 0`), knob-off verified bit-identical. Open design questions for review are in
 the morning-questions list (seed size constant, default-on policy, kinetic ground-truth case).

handoff/issue376/morning_questions.md

@@ -0,0 +1,28 @@
+# Questions for the morning
+
+1. **PR #408 review-package sharing worked?** (You flipped both; reviewers should confirm access.)
+2. **Certified-grid default**: kinetic_grid_tol is committed default-OFF (0). The B2 sweep will
+   inform a recommended value (likely 1e-3); do you want it ON by default in the kinetic example
+   decks, or opt-in only until it has survived a real DIII-D kinetic workflow?
+3. **Kinetic reference truth**: Solovev-calculated et[1] moves ~1e-3 between m64 and m256 and the
+   ladder is non-monotonic — there is no converged reference in the examples. Is there a case you
+   trust as kinetic ground truth (PENTRC benchmark?) worth wiring in before we claim accuracy wins?
+4. **G/H aliasing**: Kinetic/G vs Ideal/G/H statistics are identical in the h5 at every size —
+   expected (small corrections on a shared large scale), or worth a look?
+5. **Scope check**: certification currently applies only to kinetic_source="calculated". The
+   "fixed" test-matrix path stays on the equilibrium grid. OK?
+6. **q=5 Δ′ non-convergence** (out of scope per your earlier call) — still parked; unchanged.
+
+7. **Kinetic harness deltas vs develop are LARGE and (pending attribution) mostly inherited from
+   #398's Solovev geometry fix**: et[1] shifts of 1.7-2.5%, Im parts up to 23%, NTV torque
+   magnitude change, and — striking — **NTV ψ-quadrature evaluations 840 → 60**: the torque
+   quadrature appears to have been spending 14× the evaluations resolving geometry noise. An
+   attribution harness (vs the #398 branch instead of develop) is running/complete — see
+   attr_regress.log. Physics judgment needed: are the kinetic-case changes acceptable as
+   "the geometry-noise fix propagating into resonance-sensitive quantities" (my reading), and
+   should the kinetic harness baselines simply be re-baselined once #398 lands?
+8. **Certified-grid PR**: B1 is committed on the #408 branch (default OFF, knob-off bit-identical),
+   but arguably belongs in its own stacked PR per the plan ("kinetic work lands as its own PR
+   stacked after Phase A's"). I did NOT split it out overnight — say the word and I'll
+   cherry-pick 2225813fc onto a new branch stacked on #408 and open the PR with its review
+   package, or keep it in #408 if you prefer fewer moving pieces.