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/include/mlir/Dialect/QCO/Utils/Graph.h`:
- Around line 206-213: The adjacency bucket initialization in addDirectedEdge
uses the wrong element type, since adj_ stores Vector<IdT> but the new buckets
are created as Vector<size_t>(). Update the addDirectedEdge method in Graph to
initialize missing buckets with Vector<IdT>() so the template works for
non-size_t IdT instantiations, or otherwise explicitly constrain the graph to
size_t IDs if that is the intended design.
- Around line 186-193: The cycle-detection logic in Graph::dfs currently treats
a Seen neighbor as a back edge even when it is just the parent in an undirected
graph, so a single undirected edge can be reported as a cycle. Update the
Seen-handling branch in Graph::dfs to ignore the parent edge for
GraphType::Undirected by checking the current node against parents/top before
building and returning the cycle path, while preserving the existing behavior
for directed graphs.
- Around line 69-71: The getNodes() accessor in Graph returns an ArrayRef built
from a temporary created by to_vector(adj_.keys()), so the returned view becomes
dangling immediately. Update getNodes() to return owned storage by value (for
example a Vector/SmallVector of IdT) or otherwise ensure the node container is
stored with Graph lifetime, and adjust any callers of getNodes() to use the
owned return type rather than an ArrayRef.

In `@mlir/lib/Dialect/QCO/Transforms/Mapping/Mapping.cpp`:
- Around line 811-819: The `Mapping::` unitary handling in `Mapping.cpp` assumes
every `UnitaryOpInterface` has two indices, but single-qubit gates like `h` can
reach this path and make `indices[1]` invalid. Update the ready-item loop to
distinguish one-qubit versus two-qubit unitary ops before indexing, using the
`UnitaryOpInterface` check plus the size/arity of `indices` (or the op’s qubit
count) to route single-qubit gates through the appropriate program lookup and
only read `indices[1]` for true two-qubit cases.
- Around line 293-299: The layout refinement failure path in the mapping pass
should stop execution immediately after `signalPassFailure()` instead of falling
through to `place(...)`. Update the `generateLayout` handling in the mapping
logic so that when `failed(layout)` is true, `func->emitError()` and
`signalPassFailure()` are followed by an early return from the current
pass/function, preventing dereference of `*layout` and any work with invalid
state.
- Around line 546-585: The trial selection in generateLayout is using cumulative
swap counts from Statistics, but SABRE should rank layouts by the swaps from the
final backward pass only. Update the routing loop in generateLayout so each
Trial records the swap count from the last route<WireDirection::Backward>
invocation separately, then use that value instead of stats.nswaps when choosing
the best Trial. Keep the existing routing logic in route<WireDirection::Forward>
and route<WireDirection::Backward>, but add a dedicated per-trial field for the
final backward-pass score and compare trials against that field near the
best-trial selection.

In `@mlir/lib/Dialect/QCO/Utils/Layout.cpp`:
- Around line 37-42: The Layout::add() method currently overwrites
programToHardware_ and hardwareToProgram_ without preserving the bijection, so
add checks for any existing assignment before writing a new prog/hw pair. Update
Layout::add() to either reject duplicates or clear the previous reverse/forward
mapping for the existing program or hardware entry before assigning the new one,
ensuring programToHardware_ and hardwareToProgram_ always stay consistent.
- Around line 54-59: The public Layout::swap method currently indexes
hardwareToProgram_ directly without validating hwA and hwB, so add the same
bounds assertions used by Layout::add and the getters before reading
hardwareToProgram_ and programToHardware_. Update swap(const size_t hwA, const
size_t hwB) to check both hardware IDs are in range before performing the swaps,
keeping the validation consistent with the rest of Layout.

In `@mlir/unittests/Dialect/QCO/Transforms/Mapping/test_mapping.cpp`:
- Around line 406-409: Restore the qubits into the same tensor slots they were
extracted from by updating the qtensorInsert calls in test_mapping.cpp so the
sequence in the mapping test uses indices 0 through 3, not 1 through 4. The
issue is in the tensor reinsertion around the builder.qtensorInsert calls for
q0, q1, q2, and q3; fix the off-by-one indexing so index 0 is restored and no
write goes past the allocated size-4 tensor.

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`:
- Around line 45-46: The changelog entry currently groups `#1805` under the
initial infrastructure bullet, but it belongs with the place-and-route work.
Update the CHANGELOG.md release note so the reference to `#1805` is moved from the
initial infrastructure list to the existing place-and-route entry, keeping the
mapping aligned with the scf.for restoration support described by that PR.

In `@mlir/include/mlir/Dialect/QCO/Transforms/Mapping/Mapping.h`:
- Around line 27-29: The createMappingPass API currently takes only the coupling
edge set, which loses explicit qubit cardinality and breaks devices with
isolated qubits. Update the pass interface in Mapping.h so createMappingPass
preserves nqubits explicitly, or accepts a separate node set alongside the
DenseSet of couplings. Then propagate that preserved cardinality through the
mapping pass implementation so capacity checks and layout generation can still
account for disconnected qubits.