Measurement: the datum-only continuation case converges without Picard; the earlier finding is superseded

Underworld development team with AI support from Claude Code

Author: lmoresi

entries.md

@@ -8,7 +8,8 @@ status says whether it still stands. The topic pages hold the current reading.
 ## 2026-10
 
 - 2026-10-29 · [A per-level wall rotation cuts a rotated free-slip custom-FMG solve on an adapted annulus from 18 serial and 197 parallel velocity iterations to 3 and 4](entries/2026-10-29-per-level-wall-rotation-cuts-rotated-freeslip-fmg-iterations-on-an-adapted-annulus.md) · measurement · fault-multigrid, solvers-and-multigrid
-- 2026-10-29 · [A datum-only cold start on the rotated path sets its convergence reference from the lifted iterate and takes one increment more than it needs](entries/2026-10-29-datum-only-cold-start-reference-on-the-rotated-path-costs-one-increment.md) · finding · solvers-and-multigrid
+- 2026-10-29 · [Under continuation a datum-only cold start skips its Picard phase and still converges in 10 Newton iterations: not a must-fix](entries/2026-10-29-datum-only-continuation-skips-picard-and-still-converges.md) · measurement · solvers-and-multigrid
+- 2026-10-29 · [A datum-only cold start on the rotated path sets its convergence reference from the lifted iterate and takes one increment more than it needs](entries/2026-10-29-datum-only-cold-start-reference-on-the-rotated-path-costs-one-increment.md) · finding · solvers-and-multigrid · superseded
 - 2026-10-28 · [Which fault-mesh construction is used in practice: the placed ladder with a hand-built split chain; adapt-and-cut is the automated route the hierarchy was measured on](entries/2026-10-28-which-fault-mesh-construction-is-used-in-practice.md) · finding · fault-multigrid
 - 2026-10-28 · [The fault-split integration head reproduces UWTN 2026-022's numbers; the suites pass but for two failures that predate it](entries/2026-10-28-fault-split-integration-head-reproduces-uwtn-2026-022.md) · measurement · fault-multigrid, solvers-and-multigrid
 - 2026-10-28 · [Coarse solve on the rotated path: handle the null space, leave the coarse level to PETSc's defaults; MUMPS is not the default](entries/2026-10-28-coarse-solve-null-space-first-petsc-defaults-not-mumps.md) · decision · solvers-and-multigrid, fault-multigrid

entries/2026-10-29-datum-only-cold-start-reference-on-the-rotated-path-costs-one-increment.md

@@ -4,7 +4,8 @@ date: 2026-10-29
 authors: [wombat session_01HxWvn5rMK67zasgirdXszQ (Fault Models Code Unification), docs hub]
 options:
   kind: finding
-  status: live
+  status: superseded
+  superseded_by: 2026-10-29-datum-only-continuation-skips-picard-and-still-converges.md
   topics: [solvers-and-multigrid]
   links:
     - https://github.com/underworldcode/underworld3/tree/feature/fault-split-integration

entries/2026-10-29-datum-only-continuation-skips-picard-and-still-converges.md

@@ -0,0 +1,42 @@
+---
+title: "Under continuation a datum-only cold start skips its Picard phase and still converges in 10 Newton iterations: not a must-fix"
+date: 2026-10-29
+authors: [wombat session_01HxWvn5rMK67zasgirdXszQ (Fault Models Code Unification), docs hub]
+options:
+  kind: measurement
+  status: live
+  topics: [solvers-and-multigrid]
+  links:
+    - https://github.com/underworldcode/underworld3/issues/883
+    - https://github.com/underworldcode/underworld3/issues/876
+    - https://github.com/underworldcode/underworld3/tree/feature/fault-split-integration
+---
+
+Follows `2026-10-29-datum-only-cold-start-reference-on-the-rotated-path-costs-one-increment`,
+which left the continuation case open.
+
+## What was tried
+
+C at 4698938f: a ThinSheet on a spherical manifold (cell size 0.25), viscoplastic
+(unit viscosity, yield stress 0.05), a prescribed normal velocity $0.1 z^2$ as the only
+forcing, tolerance $10^{-6}$, `consistent_jacobian="continuation"`.
+
+## What happened
+
+The switch is evaluated at iteration 0 with a zero reference, so the Picard phase is
+skipped, and the solve converges in 10 iterations with exactly the default Newton
+tangent's residual history ($1.13\times10^{-2}$ down to $1.46\times10^{-9}$). No
+failure, no stall.
+
+## What we now think
+
+By the criterion set the day before (a must-fix only if a datum-only nonlinear solve
+stalls or fails), the hand-resolved hunk stays as it is in C. The cost on the linear
+case and the skipped Picard phase here are the same root cause, the reference taken
+from the lifted iterate, and go to the follow-up #883 (reference from the lift's
+right-hand side) as a small PR after C. Separately, `ThinSheet.solve(picard=N)` is a
+TypeError (#876), so a sheet cannot ask for Picard steps by that route either.
+
+## Verdict
+
+Kept. Closes the open question of the previous entry; #883 carries the fix.

index.md

@@ -42,6 +42,7 @@ page says what currently stands, and the weekly digest says what moved.
 ## Recent entries
 
 - 2026-10-29 · [A per-level wall rotation cuts a rotated free-slip custom-FMG solve on an adapted annulus from 18 serial and 197 parallel velocity iterations to 3 and 4](entries/2026-10-29-per-level-wall-rotation-cuts-rotated-freeslip-fmg-iterations-on-an-adapted-annulus.md) · measurement
+- 2026-10-29 · [Under continuation a datum-only cold start skips its Picard phase and still converges in 10 Newton iterations: not a must-fix](entries/2026-10-29-datum-only-continuation-skips-picard-and-still-converges.md) · measurement
 - 2026-10-29 · [A datum-only cold start on the rotated path sets its convergence reference from the lifted iterate and takes one increment more than it needs](entries/2026-10-29-datum-only-cold-start-reference-on-the-rotated-path-costs-one-increment.md) · finding
 - 2026-10-28 · [Which fault-mesh construction is used in practice: the placed ladder with a hand-built split chain; adapt-and-cut is the automated route the hierarchy was measured on](entries/2026-10-28-which-fault-mesh-construction-is-used-in-practice.md) · finding
 - 2026-10-28 · [The fault-split integration head reproduces UWTN 2026-022's numbers; the suites pass but for two failures that predate it](entries/2026-10-28-fault-split-integration-head-reproduces-uwtn-2026-022.md) · measurement
@@ -60,6 +61,5 @@ page says what currently stands, and the weekly digest says what moved.
 - 2026-10-17 · [Box fault: implementation status (mesher, damage band, ported fixes)](entries/2026-10-17-box-fault-implementation-status-mesher-damage-band-ported-fixes.md) · finding
 - 2026-10-17 · [Box fault: implementation as built (gmsh patch + registered coarse tail, sign, cost)](entries/2026-10-17-box-fault-implementation-as-built-gmsh-patch-registered-coarse.md) · measurement
 - 2026-10-17 · [Box fault: first results - the x-parallel null and the oblique walk](entries/2026-10-17-box-fault-first-results-the-x-parallel-null-and-the-oblique-walk.md) · finding
-- 2026-10-17 · [Box fault: ETD expression-viscosity bug, stress-history flavour, cost](entries/2026-10-17-box-fault-etd-expression-viscosity-bug-stress-history-flavour.md) · measurement
 
 [All entries](entries.md)

myst.yml

@@ -50,6 +50,7 @@ project:
         - title: "2026-10"
           children:
             - file: entries/2026-10-29-per-level-wall-rotation-cuts-rotated-freeslip-fmg-iterations-on-an-adapted-annulus.md
+            - file: entries/2026-10-29-datum-only-continuation-skips-picard-and-still-converges.md
             - file: entries/2026-10-29-datum-only-cold-start-reference-on-the-rotated-path-costs-one-increment.md
             - file: entries/2026-10-28-which-fault-mesh-construction-is-used-in-practice.md
             - file: entries/2026-10-28-fault-split-integration-head-reproduces-uwtn-2026-022.md

topics/solvers-and-multigrid.md

@@ -97,7 +97,7 @@ one SNES solve, $\alpha$ set per iteration, monotone except for a back-off when
 ## Entries
 
 - 2026-10-29 · [A per-level wall rotation cuts a rotated free-slip custom-FMG solve on an adapted annulus from 18 serial and 197 parallel velocity iterations to 3 and 4](../entries/2026-10-29-per-level-wall-rotation-cuts-rotated-freeslip-fmg-iterations-on-an-adapted-annulus.md) · measurement
-- 2026-10-29 · [A datum-only cold start on the rotated path sets its convergence reference from the lifted iterate and takes one increment more than it needs](../entries/2026-10-29-datum-only-cold-start-reference-on-the-rotated-path-costs-one-increment.md) · finding
+- 2026-10-29 · [Under continuation a datum-only cold start skips its Picard phase and still converges in 10 Newton iterations: not a must-fix](../entries/2026-10-29-datum-only-continuation-skips-picard-and-still-converges.md) · measurement
 - 2026-10-28 · [The fault-split integration head reproduces UWTN 2026-022's numbers; the suites pass but for two failures that predate it](../entries/2026-10-28-fault-split-integration-head-reproduces-uwtn-2026-022.md) · measurement
 - 2026-10-28 · [Coarse solve on the rotated path: handle the null space, leave the coarse level to PETSc's defaults; MUMPS is not the default](../entries/2026-10-28-coarse-solve-null-space-first-petsc-defaults-not-mumps.md) · decision
 - 2026-10-16 · [Dead end: Preconditioning an indefinite viscoelastic velocity block](../entries/2026-10-16-dead-end-preconditioning-an-indefinite-viscoelastic.md) · dead-end
@@ -116,4 +116,8 @@ one SNES solve, $\alpha$ set per iteration, monotone except for a back-off when
 - 2026-10-05 · [Dead end: Ladders in V and in the rate-weakening cohesion homotopy](../entries/2026-10-05-dead-end-ladders-in-v-and-in-the-rate-weakening-cohesion.md) · dead-end
 - 2026-10-05 · [Dead end: A fixed-h `-snes_test_jacobian`](../entries/2026-10-05-dead-end-a-fixed-h-snes-test-jacobian.md) · dead-end
 
+Retired (superseded, absorbed or discarded):
+
+- 2026-10-29 · [A datum-only cold start on the rotated path sets its convergence reference from the lifted iterate and takes one increment more than it needs](../entries/2026-10-29-datum-only-cold-start-reference-on-the-rotated-path-costs-one-increment.md) · finding · superseded → 2026-10-29-datum-only-continuation-skips-picard-and-still-converges.md
+
 <!-- END ENTRIES -->