Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@CHANGELOG.md`:
- Line 14: Replace the incorrect changelog bullet "✨ Add Dead Gate Elimination
Pattern ([`#1755`]) ([**`@DRovara`**])" with an entry that references this PR
(`#1756`) and accurately describes the implemented feature (measurement lifting /
classical-control replacement), e.g., update the PR number to 1756 and change
the description to mention "Measurement lifting and classical-control
replacement" so the released notes match the actual code changes.

In `@mlir/include/mlir/Dialect/QCO/QCOUtils.h`:
- Around line 240-276: The header's inline function checkAndRemoveDeadGate uses
QCO-specific types (SinkOp, UnitaryOpInterface, MeasureOp, IfOp, Reset) but
doesn't include their declarations, making the header order-dependent; fix by
making the header self-contained: add a direct include of the QCO ops/declaring
header (e.g., include "mlir/Dialect/QCO/QCOOps.h") at the top of
mlir/include/mlir/Dialect/QCO/QCOUtils.h so dyn_cast/isa and method calls
resolve, or alternatively move the implementation of checkAndRemoveDeadGate out
of the header into a .cpp that includes QCOOps.h if you prefer to avoid
additional header dependencies.

In `@mlir/lib/Dialect/QCO/Transforms/Optimizations/ReplaceClassicalControls.cpp`:
- Around line 143-155: The lambda passed to IfOp::create currently always builds
a CtrlOp (via CtrlOp::create) even when newControls (derived from toKeep.size())
is empty, which yields invalid IR because qco::CtrlOp requires >=1 control;
modify the lambda to branch on newControls.empty(): if non-empty, keep the
existing flow (create CtrlOp, inline newCtrl.getRegion() into op.getRegion(),
return newCtrl.getOutputQubits()); if empty, instead directly clone or move the
original unitary body into the IfOp region (use rewriter to inline/clone the
original region/body into the new IfOp region) and return the appropriate output
qubits, ensuring no zero-control CtrlOp is constructed (affects
ReplaceClassicalControls.cpp paths like replaceClassicalControlsOnlyControl and
replaceClassicalControlsTwoOfTwoControls).
- Around line 58-63: The isDiagonal predicate omits RzOp so controlled-rz paths
are not recognized; update the isDiagonal function (the dyn_cast<InvOp> branch
and the isa<> return) to include RzOp among the recognized diagonal ops (i.e.,
add RzOp to the isa<ZOp, SOp, TOp, POp, SdgOp, TdgOp, IdOp> list) so
controlled-`rz` will take the diagonal-swap path.

In
`@mlir/unittests/Dialect/QCO/Transforms/Optimizations/test_qco_replace_classical_controls.cpp`:
- Around line 377-380: Remove the leftover debug printing by deleting the
module->dump() call at the end of the test in
test_qco_replace_classical_controls.cpp; keep the existing assertions
runReplaceClassicalControlsPass(module.get()).succeeded() and
runCanonicalizerPass(reference.get()).succeeded() but remove the module->dump()
line so the test no longer prints the transformed module to CI logs.

---

Outside diff comments:
In `@mlir/include/mlir/Dialect/QCO/IR/QCOOps.td`:
- Around line 1071-1077: The QCO region wrapper ops declare the Pure trait which
expands to AlwaysSpeculatable + NoMemoryEffect and incorrectly
duplicates/overrides the region-modelled effects; remove the Pure trait from the
op definitions for CtrlOp, InvOp, and IfOp (the TD defs named CtrlOp, InvOp,
IfOp) so they only keep RecursiveMemoryEffects (and other existing traits like
UnitaryOpInterface, AttrSizedOperandSegments, AttrSizedResultSegments,
SameOperandsAndResultType, SameOperandsAndResultShape,
SingleBlockImplicitTerminator) to ensure effects are solely driven by the nested
region.

In `@mlir/lib/Dialect/QCO/Builder/QCOProgramBuilder.cpp`:
- Around line 1105-1165: The finalize paths currently emit a func::ReturnOp
without validating or matching the main function's declared result types
(initialize(TypeRange) may set a non-i64 or different arity), so update
QCOProgramBuilder::finalize() and QCOProgramBuilder::finalize(ValueRange) to
first query mainFunc.getType().getResults(), assert that returnValues.size()
equals that result list size and that each returnValues[i].getType() ==
expectedResults[i], and if the no-arg finalize() is used produce appropriate
default constants per expectedResults type (not always an i64 zero) instead of
unconditionally calling intConstant(0); then create the func::ReturnOp with
validated/matched values (and emit a clear error via llvm::reportFatalUsageError
if arity/types mismatch). Ensure references: finalize(), finalize(ValueRange),
initialize(TypeRange), func::ReturnOp::create, mainFunc.getType().getResults(),
and intConstant are used when implementing the checks and per-type default
constant construction.

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In
`@mlir/unittests/Dialect/QCO/Transforms/Optimizations/test_qco_measurement_lifting.cpp`:
- Around line 534-574: Correct the expected module in
dontLiftMeasurementMultipleGatesInControl by keeping the control region before
measuring its control qubit, matching the transformed program’s operation order.
Update the referenceBuilder sequence so r0S0 is passed through ctrl first, then
measure the resulting control state, while preserving the multi-gate z/x control
body and final comparison.

In
`@mlir/unittests/Dialect/QCO/Transforms/Optimizations/test_qco_replace_classical_controls.cpp`:
- Line 137: Remove the const qualifier from the ValueRange parameters in all
five lambdas, including those in the QCO optimization test locations identified
by the symbols or surrounding lambda signatures. Keep the parameter types as
ValueRange and change only the signature qualifiers.