Issue quick-fix batch 2: six small independent fixes (#415 #447 #413 #411 #451 #414)

[lmoresi]

Summary

Six small, independent fixes, each with fail-first evidence where demonstrable. Closes #415, closes #447, closes #413, closes #411, closes #451, closes #414 (manual close on merge — PRs target development).

Issue	Fix
#415	UWexpression.is_comparable → False (plain-Symbol semantics). sympy.Max/Min(expr, number) previously crashed on Float internals (_prec); now constructs symbolically, resolves on unwrap, stays live to constant ramps. Pre-fix AttributeError reproduced.
#447	Parameter assignment guards both self-reference cycles: read-back write is a no-op (was: container nested in itself → RecursionError on .value); composite P.x = P.x * 2 snapshots current contents. Pre-fix RecursionError reproduced.
#413	2D boundary-flux recovery detects trace order from the reaction data and assembles the matching P1 two-node / P2 three-node line mass (lumped + consistent); mixed traces raise legibly. Was a bare KeyError on P1.
#411	compute_volume_residual_fields + compute_boundary_residual_fields get the #407 DMPlexInsertBoundaryValues insert (with residual_time plumbing). Regression test pins exact agreement with _assemble_volume_reaction and fails on the pre-fix build.
#451	Spiegelman notch example builds gmsh geometry in ND units (S = 1.0) — was 30× too large in every direction.
#414	Curved-boundary warning (P2 vertices slow, midpoints superconvergent, with measured numbers) on the three consumer docstrings.

Validation

• New: tests/test_0104_expression_parameter_guards.py (3 tests); test_1019_boundary_flux.py +3 (P1 trace lumped+consistent, compute_volume_residual_fields / boundary sibling still strip Dirichlet values (the #407 fix was applied to one of three duplicated variants) #411 exact-agreement).
• Fail-first: sympy.Max/Min(UWexpression, plain number) crashes at construction: AttributeError '_prec' #415/Self-assigning a constitutive Parameter (read then write back) recurses to stack death #447 repros crash on the pre-fix build; the compute_volume_residual_fields / boundary sibling still strip Dirichlet values (the #407 fix was applied to one of three duplicated variants) #411 test FAILS on the pre-fix build and passes here.
• Worktree gate level_1 and tier_a: 523 passed, 0 failed (17 skipped, 1 xfail).

Adversarial review to follow on this thread before merge, per project policy.

Underworld development team with AI support from Claude Code

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[lmoresi]

Adversarial review — live probes on the batch worktree

Verdict: APPROVE with one scope caveat, addressed in a response commit. All six changes hold under attack. The caveat: the #447 guard as submitted covers the two direct cycles but not a self-reference nested inside a wrapper expression — and that hole's worst mode is silent garbage, not a crash. Hardened in the response commit (recursive cycle detection → legible ValueError), so #447 closes fully. One pre-existing (not-a-regression) P3 defect discovered en route is filed separately.

Findings

MAJOR (scope) — deeper #447 cycle was unguarded — PROBED, fixed in response commit. wrapper = uw.expression("w", Params.yield_stress * 2); Params.yield_stress = wrapper was accepted; afterwards constitutive_model.flux → RecursionError (the original crash site) and unwrap_expression(...) → silently returns 2⁵⁰·τ_y — the cycle unrolled ~50 times against an internal depth cap and handed back as a live expression. The response commit adds recursive containment detection through nested UWexpression contents (visited-set bounded): any assignment that would create a cycle through wrappers now raises a ValueError telling the user to snapshot explicitly. Cross-container cycles (a→b→a) are caught by the same walk.

MINOR — Matrix-valued composite snapshot produced malformed sympy — PROBED, fixed in response commit. xreplace of a Matrix-valued container inside a Matrix entry stored nested-matrix garbage silently. Now raises for Matrix-shaped contents instead of substituting.

NOTE — pre-existing P3 silent-wrong boundary flux — PROBED both sides, filed as a separate issue. A degree-3 trace returns finite but wrong flux (−0.57…−0.74 vs exact −1.0) — byte-identical on origin/development: both P3 edge DOFs collapse onto the midpoint coordinate key and one reaction overwrites the other, on the old code and the new. Not a regression from this batch; the recovery should reject degree > 2 traces explicitly.

NOTE — doc touch-ups, fixed in response commit: stale "km" comment in the Spiegelman example (line ~123) contradicting the corrected framing; "Flat boundaries are pointwise exact" softened to "exact up to solver tolerance" (the tests assert 1e-3).

Attacks that failed (evidence)

• sympy.Max/Min(UWexpression, plain number) crashes at construction: AttributeError '_prec' #415 blast radius (the big risk): Poisson with diffusivity = sympy.Max(k, 0.5) SOLVED and ramped k = 1.0 → 0.1 → 2.0 across the threshold without rebuild: T_max = 0.125001 / 0.250001 / 0.062501 vs analytic 1/(8·max(k,0.5)), matching fresh solvers to ≤3e-6. Piecewise conditions resolve live; default_sort_key/hash fine; full srepr byte-identical across two fresh processes (parallel JIT codegen determinism holds); uw.function.evaluate of Max(k, ·) correct before and after ramp.
• Self-assigning a constitutive Parameter (read then write back) recurses to stack death #447 guards that ARE in the submitted diff: read-back no-op skips _reset correctly while all other branches still reset (instance-level spy); cross-parameter alias (P.a = P.b) stays a live reference (no false no-op — 2.5 → 4.0 after updating the source); lazy mixed composite P.x = P.x*2 + P.y*3 gives exactly the intended semantics (self snapshotted, other container live).
• boundary_flux: 2D P1 traces crash with a bare KeyError in the lumped line-mass assembly #413/compute_volume_residual_fields / boundary sibling still strip Dirichlet values (the #407 fix was applied to one of three duplicated variants) #411 at np2: test_1019 8 passed (both new P1 mass modes + the compute_volume_residual_fields / boundary sibling still strip Dirichlet values (the #407 fix was applied to one of three duplicated variants) #411 exact-agreement test hold per-rank); parallel boundary-flux suite 3 passed. time=0.5 path matches _assemble_volume_reaction(time=0.5) to rtol 1e-12; Nitsche boundary-residual path runs clean with the new insert.
• Spiegelman notch benchmark example builds a 3600 x 900 km domain instead of 120 x 30 km #451 reading sweep: every geometry entity scales with S (arcs, centres, mesh sizes); no residual km assumptions (viz windows derive from mesh bounds; the symbolic lithostatic pressure rescales with the frame — and was itself 30× wrong before this fix, consistently repaired).
• Mixed-trace raise is rank-symmetric by construction (allgathered inputs).

Adversarial review per project policy — Underworld development team with AI support from Claude Code