Nice optimization. Filter pull-up above aggregates, maps, integrate/differentiate, noop, and filter-merge. Read carefully; the safety guards look right:

• Fanout=1 gate correctly prevents duplicating the source when it has other consumers.
• Aggregate case uses NoExpression in the value slot and ContainsNoExpression.search after reduction to reject filters that actually depend on the aggregate's output — a clean way to prove key-only dependence.
• Map/MapIndex relies on shouldInlineComposition — same threshold used elsewhere, so complexity blow-up stays bounded.
• Integrate/Differentiate/Noop commute with a stateless filter on Z-sets, so pulling through is sound.
• Filter+Filter merge into AND and calls .after(...) on the source's RelNode to preserve provenance — good.

DBSPCompiler.getFinalCircuit guard if (circuit != null) is a sensible fix for the failed-compilation path.

Persistent-id churn in metadataTests-generateDFRecursive.json is expected (circuit shape changed) and lives only in a compiler test fixture — no runtime checkpoint compatibility concern.

Two nitpicks, non-blocking:

• The filter-merge branch calls .copyAnnotations(operator) on the combined filter but drops annotations from the upstream filter (source). Probably fine for now, but if either filter carried something meaningful (e.g. hoisted-from-view markers), it's gone.
• Utilities.showResultSet signature reformat is unrelated to the optimization — trivial, keep it, just noting.

LGTM.

Re-review at tip 67e0722 (delta since 39cb149).

Delta is cleanly additive:

• PullFilterVisitor: new branch handling DBSPConstantOperator sources. For each (row, weight) in the constant Z-set, calls the filter closure, reduces and Simplifys the result; if every element folds to a DBSPBoolLiteral non-null value, replaces filter(const) with a filtered constant. Bails conservatively on NULL results or anything that doesn't collapse to a bool literal. Sound — filter is pure and inputs are literals.
• CircuitOptimizer: adds PropagateEmptySources after PullFilterVisitor (natural follow-up now that filters can produce empty constants).
• SetOpTests.simplifyIntersect: verifies INTERSECT with an unmatchable row (5, NULL against a NOT NULL column) collapses to a constant/source/sink — good end-to-end coverage of the new path.

Nits (still non-blocking, same shape as before):

• Filter annotations aren't copied onto the new DBSPConstantOperator. Probably fine since it's a constant, but consistent with the earlier nit.
• The "Should not happen" comment for b.value == null is a bit optimistic — WHERE semantics treat NULL as false, so folding null-bool to drop-the-row would recover a few more cases. Conservative bail is safe, just leaves optimization on the table.

Approving.

Re-review at 82714218 — APPROVE.

Rebase onto main plus a small cleanup: the inline DBSPConstantOperator branch in PullFilterVisitor.postorder(DBSPFilterOperator) has been extracted into PropagateConstants.filterConstant(compiler, constOp, filterClosure). Same semantics as before (per-row reduce + Simplify, bail if any result is non-DBSPBoolLiteral or null), just with the loop lifted into a shared helper — less duplication and easier to reason about. PullFilterVisitor now only imports what it still uses; the HashMap/DBSPZSetExpression/DBSPBoolLiteral/Simplify imports are dropped, which is nice.

No behavioural change from the version I approved at 67e0722f.