[ChrisRackauckas-Claude]

Summary

Stacked on #59 (merge after it). ≲/≳ build Symbolics.Inequality constraint objects, not Boolean comparisons; Symbolics 7 leaving them as unfolded isless(...) inside evaluate is a design difference, not an upstream bug. This PR maps Inequality(lhs, rhs, leq|geq) to lhs <= rhs / lhs >= rhs via boolean_probability_event before probability-oracle evaluation (canonical moments and Winkler paths), so Flood-style events become Bool.

Verification

Fail-before / pass-after

Before (#59 tip sources, new tests):

• Unit: UndefVarError: boolean_probability_event not defined
• Flood test_canonical_moments.jl:29: MethodError: isapprox(::BasicSymbolic, ::Float64)

After:

Test Summary: | Pass  Total     Time
smoke         |   32     32  1m03.7s

(includes unit conversion + Flood canonical-moments objective)

OUQBase Core / QA

julia +lts (1.10.12):

promote_symtype operators |   22     22
Flood Problem (canonical moments objective) |    1      1
Testing OUQBase tests passed
Quality Assurance |   16     16
Testing OUQBase tests passed

julia +1 (1.13.0):

promote_symtype operators |   22     22
Flood Problem (canonical moments objective) |    1      1
Testing OUQBase tests passed
Quality Assurance |   18     18
Testing OUQBase tests passed

What I did NOT verify

• End-to-end optimizer solves beyond the Flood objective smoke at the fixture point.
• Root-package / Downgrade workflows.

Reviewer pushback

• Whether converting at extract time (current) vs inside the per-sample evaluate closure is clearer.
• Whether any other probability path still evaluates a raw Inequality.

Part of #54. Stacked on #59 — merge after #59.

Please ignore this draft until reviewed by @ChrisRackauckas.

Risk assessment

• Risk: low–medium
• Blast radius: probability-oracle evaluation for ≲/≳ events only (canonical moments + Winkler Symbolic).
• Evidence: fail-before Flood isapprox(BasicSymbolic,…) / pass-after numeric; Core+QA green on +lts and +1.
• Independent review: Codex gpt-6-astra round 1 CHANGES (non-public Symbolics.leq/geq, fixed); Claude Code claude-opus-5-5 (a different harness from the Cursor author) round 2 on head 133a780: VERDICT MERGE, no blocking findings. Merge after Fix OUQBase 𝔼/ℙ SymbolicUtils Operator ambiguity (separate from #54 floor) #59.
• Merge: after Fix OUQBase 𝔼/ℙ SymbolicUtils Operator ambiguity (separate from #54 floor) #59 + human review

Harness: Cursor Agent CLI 2026.09.26-dd393fe; model: auto; transcript: /home/crackauc/sandbox/agent-jobs/OptimalUncertaintyQuantification.jl/jobs/54-cursor/log.txt on amdci2.julia.csail.mit.edu

Update (Codex gpt-6-astra round-1 CHANGES)

Stopped using non-public Symbolics.leq / Symbolics.geq. Inequality dispatch now compares relational_op against the values from the public constructors (lhs ≲ rhs).relational_op / (lhs ≳ rhs).relational_op (also applied in the Winkler Jump helper). Removed :leq / :geq from the QA qualified-accesses ignore list. Rebased onto fixed #59 (ModelingToolkitBase.getdefault, no ownership ignore).

Re-verification (OPTIMALUNCERTAINTYQUANTIFICATION_TEST_GROUP)

julia +lts (1.10.12):

promote_symtype operators |   31     31
Flood Problem (canonical moments objective) |    1      1
Testing OUQBase tests passed
Quality Assurance |   16     16
Testing OUQBase tests passed

julia +1 (1.13.0):

promote_symtype operators |   31     31
Flood Problem (canonical moments objective) |    1      1
Testing OUQBase tests passed
Quality Assurance |   18     18
Testing OUQBase tests passed

What I did NOT verify

• End-to-end Winkler optimizer construction (pre-existing getbounds(::OrderedSet) failure on Fix OUQBase 𝔼/ℙ SymbolicUtils Operator ambiguity (separate from #54 floor) #59 tip and this head; not attributed to this conversion).
• Root-package / Downgrade workflows.

Harness: Cursor Agent CLI 2026.09.26-dd393fe; model: auto; transcript: /home/crackauc/sandbox/agent-jobs/OUQ-b/jobs/54-cursor/log.txt on amdci2.julia.csail.mit.edu

Risk assessment

• Risk: medium
• Blast radius: OUQBase only. Specifically: how the 𝔼/ℙ/𝟙 operator terms are built, random-variable group resolution (get_ordered_group_names now looks inside 𝔼/ℙ and inside ≲/≳ inequalities), raw-moment order parsing, and probability-objective evaluation in the canonical-moments oracle path and the Winkler Symbolic path. Winkler's JuMP path is unaffected because it still reads the raw condition (only(arguments(objective._obj))). The PR's base is main, not Fix OUQBase 𝔼/ℙ SymbolicUtils Operator ambiguity (separate from #54 floor) #59, so merging it also brings in all 8 of Fix OUQBase 𝔼/ℙ SymbolicUtils Operator ambiguity (separate from #54 floor) #59's commits. That includes a new direct dependency on ModelingToolkitBase (compat 1.50).
• Evidence: On head 133a780, 2 checks fail. Both are pre-existing, because the same jobs fail with the same errors on the latest main push runs (013f14a, 2026-09-16):
  • Downgrade Sublibraries / test (lib/DiscreteMeasures): 5 errors, ArgumentError: intersect is purposely not supported for intervals at discrete_measures_tests.jl:94-98. Main shows the same 5 errors on the same lines. Pre-existing, and the PR does not touch lib/DiscreteMeasures.
  • tests / QA (julia 1) / Tests - QA (Julia 1): the root package's no_implicit_imports errors, and public API has docstrings and No unapproved public reexports fail. The totals are 18 passed / 2 failed / 1 errored, which matches main exactly. Pre-existing.
  • Every other check passes, including OUQBase Core on lts/1/pre, OUQBase QA, root Core ×4, Docs, Runic and typos. Downgrade Sublibraries / test (lib/OUQBase) fails on main and passes here, so that one is an improvement.
  • No regressions. One caveat: Fix OUQBase 𝔼/ℙ SymbolicUtils Operator ambiguity (separate from #54 floor) #59's own head (630a206) is red on sublibraries / lib/OUQBase / Julia {1,lts} / Tests - Core with MethodError: isapprox(::BasicSymbolic, ::Float64) at test_canonical_moments.jl:29. Following the description's advice to merge Fix OUQBase 𝔼/ℙ SymbolicUtils Operator ambiguity (separate from #54 floor) #59 first would leave main red on OUQBase Core until this PR lands.
• Independent review: Devin CLI (Mac) / fusion-claude-opus-5-5-high-sidekick-swe-2-medium rated it medium, medium confidence: CI shows no regressions and the new tests fail before the fix and pass after. However, this is a stacked PR that folds in all of Fix OUQBase 𝔼/ℙ SymbolicUtils Operator ambiguity (separate from #54 floor) #59's symbolic-internals rework, and the Winkler Symbolic change has no end-to-end test (the author says a pre-existing getbounds(::OrderedSet) failure blocks one).
• Merge: needs human review. Findings:
  1. Stacked PR. The diff contains all of Fix OUQBase 𝔼/ℙ SymbolicUtils Operator ambiguity (separate from #54 floor) #59, which is still open and red on its own. The advised order ("merge after Fix OUQBase 𝔼/ℙ SymbolicUtils Operator ambiguity (separate from #54 floor) #59") would leave main temporarily red. Merging this PR alone, or retargeting it onto Fix OUQBase 𝔼/ℙ SymbolicUtils Operator ambiguity (separate from #54 floor) #59's branch, avoids that.
  2. Undisclosed new dependency. The new direct dependency ModelingToolkitBase (compat 1.50, SciML/MIT) is not mentioned or license-noted in this PR's description. OUQBase's version is not bumped either.
  3. Internal field access remains. .relational_op is still read directly from Symbolics.Inequality. Comparing it against the value from the public ≲/≳ constructors avoids naming Symbolics.leq/geq, but it still depends on an undocumented struct field. The QA ignores for promote_symtype, isbinop, evaluate, BasicSymbolic, Operator, Term, symtype and value are still in place (pre-existing).
  4. Behavior changes in exported code, no SemVer bump. Exported 𝔼/ℙ/𝟙 now build terms with SymbolicUtils.term instead of Term{symtype(x)}. get_raw_moment_order now errors on any left-hand side that isn't a 𝔼 term, and build_raw_moment_sequence converts the right-hand side to Float64. These look like bug fixes and all touched helpers are unexported, but a human should confirm none of it counts as a public behavior change.
  5. Tests are strengthened, not weakened. The QA ignore list got shorter (:geq, :leq, :getdefault removed) and there are no skips or loosened tolerances. Two small issues in the new test file: d = Dict(...) is unused, and one assertion checks that Symbolics leaves isless unfolded, which will break if Symbolics changes that behavior.
  6. Not verified by the author: end-to-end optimizer solves, the Winkler optimizer build, and root/Downgrade workflows (all stated in the PR description). The PR has no human review comments.

---

🤖 Risk assessment posted by an AI agent (fleet master) — harness: Devin CLI (Mac) / fusion-claude-opus-5-5-high-sidekick-swe-2-medium; dispatched by Claude Code head, model claude-opus-5-5[1m]
Conversation: local Claude Code session 3cd6500a-1f81-46b5-ac0b-c466e15b6a53 on Chris's Mac (session ID, no URL)