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 `@src/qdmi/driver/CMakeLists.txt`:
- Around line 55-60: The rpath entries are currently set with
target_link_options(... INTERFACE ...), which only propagates to consumers and
not to the driver's own binary (so dlopen() inside the driver won't find device
libs); replace that with set_target_properties(${TARGET_NAME} PROPERTIES
BUILD_RPATH "<list>" (or INSTALL_RPATH if you want it for install) using the
same generator expressions $<TARGET_FILE_DIR:MQT::CoreQDMINaDevice>,
$<TARGET_FILE_DIR:MQT::CoreQDMIScDevice>,
$<TARGET_FILE_DIR:MQT::CoreQDMI_DDSIM_Device> so the rpath is embedded into the
driver itself; remove the INTERFACE target_link_options entries or keep only
non-INTERFACE options if needed.

In `@src/qdmi/driver/Driver.cpp`:
- Around line 44-81: Add Doxygen-style comments for the two Windows loader
helpers so their behavior is clear; above getDriverDirectory() document what it
does (returns the directory of the current module), note it is [[nodiscard]],
describe failure behavior (returns empty path), and summarize the Windows APIs
used (GetModuleHandleExW/GetModuleFileNameW) and why a loop/resizing is
necessary; above loadDeviceLibrary(const std::string& libName) document the
parameter (libName), explain path resolution (if libName has a parent path use
it, otherwise resolve relative to getDriverDirectory()), describe the return
value (HMODULE or null on failure), and list that it uses LoadLibraryExW with
LOAD_LIBRARY_SEARCH_DLL_LOAD_DIR | LOAD_LIBRARY_SEARCH_DEFAULT_DIRS; include
`@param`, `@return` and brief one-line description for each function following the
project Doxygen style.
- Around line 383-386: Driver::Driver() currently calls addDynamicDeviceLibrary
for each entry in DYN_DEV_LIBS and allows any dlopen/LoadLibrary exception to
bubble out (breaking Driver::get() and C entrypoints). Wrap each
addDynamicDeviceLibrary(lib, prefix) invocation in a try/catch that catches
std::exception (and a catch(...) as fallback), log a clear error including the
failing lib/prefix and the exception.what() (or a generic message for non-std
exceptions), and continue to the next library so missing builtin DLLs are
skipped instead of aborting startup.

In `@test/python/plugins/qiskit/test_backend.py`:
- Around line 68-92: Add Google-style docstrings to the four helper functions
_single_qubit_circuit, _two_qubit_circuit, _two_qubit_barrier_circuit, and
_two_qubit_circuit_without_measurements: for each function include a one-line
summary, a short description of what the circuit contains, a Args section (none
or describe omitted params if any), and a Returns section specifying that a
QuantumCircuit is returned; place the docstring immediately below the def line
in triple quotes following Google-style formatting.

In `@test/python/plugins/qiskit/test_estimator.py`:
- Around line 142-143: Update the stale test comment in test_estimator.py:
replace "mock random results" with text stating the test runs against the DDSIM
simulator (deterministic up to sampling noise) and mention that for the
Bell-state circuit the expected expectation values are known (II=1, XX=1, YY=-1,
ZZ=1, IX=0, YI=0) and can be asserted with pytest.approx; optionally add those
pytest.approx assertions against the estimator outputs to make the test
deterministic instead of relying on a no-crash check.
- Around line 24-33: The estimator fixture duplicates DDSIM device discovery
logic found in test_sampler.py; extract that discovery into a shared fixture
named ddsim_backend in test/python/plugins/qiskit/conftest.py (use the same
logic that creates QDMIBackend(device=device, provider=None) from fomac.Session
devices) and then change def estimator() to def estimator(ddsim_backend) ->
QDMIEstimator: return QDMIEstimator(ddsim_backend), replacing the repeated
fomac.Session/QDMIBackend code; reference the existing estimator fixture,
QDMIEstimator, QDMIBackend, and the sampler fixture in test_sampler.py when
implementing the shared ddsim_backend fixture.

In `@test/python/plugins/qiskit/test_mock_backend.py`:
- Around line 317-378: Tests repeat the same mock_get_devices closure and
monkeypatch.setattr(fomac.Session, "get_devices", ...) in multiple tests;
extract this into a small helper (e.g., _patch_session_devices) that accepts
monkeypatch and a list of MockQDMIDevice and installs a single closure returning
that list via monkeypatch.setattr, then replace the per-test mock_get_devices
definitions and monkeypatch.setattr calls with calls to the new helper in
test_backend_warns_on_unmappable_operation,
test_backend_warns_on_missing_measurement_operation, and
test_backend_validation_uses_inverse_mapping.

In `@test/python/plugins/qiskit/test_sampler.py`:
- Around line 24-33: Extract the DDSIM discovery loop into a shared pytest
fixture (e.g., ddsim_backend) in test/python/plugins/qiskit/conftest.py that
creates and returns a QDMIBackend for the device whose device.name() contains
"DDSIM" or calls pytest.skip("DDSIM device not available") if none found; then
update the sampler fixture in test_sampler.py to accept that ddsim_backend
fixture and return QDMISampler(ddsim_backend) instead of repeating the loop, and
likewise change the estimator fixture (and the similar code in test_backend.py)
to depend on ddsim_backend so all device selection and skip behavior is
centralized.

---

Outside diff comments:
In `@test/python/plugins/qiskit/test_estimator.py`:
- Around line 36-51: The test_estimator_run_simple_observable has a stale
comment about a "mock backend" and only asserts shapes; update it to assert the
actual expected expectation values from the deterministic DDSIM run: remove or
update the comment and add assertions on pub_result.data["evs"] using
pytest.approx to check evs[0] == approx(1.0) and evs[1] == approx(0.0) (and keep
the shape asserts if desired); reference the test function
test_estimator_run_simple_observable, the estimator.run(...) call and the
pub_result.data["evs"] / ["stds"] fields when making the assertions.

In `@test/python/plugins/qiskit/test_sampler.py`:
- Around line 56-58: The comment is stale because the test now runs on a real
DDSIM backend; update the test around bit_array.get_counts() by removing the
"mock backend" phrasing and add an assertion that the observed outcome keys are
only "00" and "11" (e.g., assert set(counts.keys()) <= {"00","11"}) in addition
to the existing total-count check so the test verifies the expected
deterministic Bell-state outcomes; locate this change in the test function using
bit_array.get_counts() and the counts variable.

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 `@pyproject.toml`:
- Around line 52-53: The docs dependency group is allowing
qiskit[qasm3-import]>=1.0.0 which mismatches the other groups; update the
`dependency-groups.docs` entry for `qiskit[qasm3-import]` to use the same lower
bound `>=1.1.0` so it matches `project.optional-dependencies.qiskit` and
`dependency-groups.test` (edit the docs group line that currently mentions
`qiskit[qasm3-import]` and change the version specifier to `>=1.1.0`).

In `@test/na/fomac/CMakeLists.txt`:
- Around line 18-35: Change the add_custom_command invocations that copy device
DLLs to use POST_BUILD instead of PRE_BUILD and add per-target existence guards
before referencing each device target; specifically, wrap the copy command for
each device target (MQT::CoreQDMIScDevice, MQT::CoreQDMINaDevice,
MQT::CoreQDMI_DDSIM_Device) in an if(TARGET <device>) ... endif() block and
replace PRE_BUILD with POST_BUILD while keeping the same COMMAND and destination
using $<TARGET_FILE_DIR:${TARGET_NAME}> so the commands only run when the device
targets exist and run after the test target is built.

In `@test/python/plugins/qiskit/test_mock_backend.py`:
- Around line 176-178: The outcome generation incorrectly encodes zero classical
bits because format(0, "00b") yields "0"; update the logic in the block that
computes num_outcomes and outcomes (using self._num_clbits and the outcomes
list) to special-case self._num_clbits == 0 so that outcomes becomes [""] (and
num_outcomes remains 1) instead of producing a phantom "0" bit string; adjust
only the outcome generation around num_outcomes and outcomes to return an
empty-string key for no-measurement programs.

In `@test/qdmi/devices/dd/CMakeLists.txt`:
- Around line 36-43: The custom command that copies the Windows DLL uses
PRE_BUILD which is unreliable across non-Visual-Studio generators; change the
add_custom_command for TARGET ${TARGET_NAME} to use POST_BUILD instead of
PRE_BUILD so the copy runs after the target is built (staging
$<TARGET_FILE:MQT::CoreQDMI_DDSIM_Device> into
$<TARGET_FILE_DIR:${TARGET_NAME}>). Ensure the command still uses the same
TARGET ${TARGET_NAME} and copy_if_different arguments but with POST_BUILD to
avoid timing issues with Ninja and other generators.

In `@test/qdmi/driver/CMakeLists.txt`:
- Around line 19-36: The add_custom_command calls for the test target (TARGET
${TARGET_NAME}) use PRE_BUILD which is generator-dependent; change each
PRE_BUILD to POST_BUILD and make the copy commands depend explicitly on the test
target's build so they run after the target is produced, i.e. update the three
add_custom_command invocations that reference
$<TARGET_FILE:MQT::CoreQDMIScDevice>, $<TARGET_FILE:MQT::CoreQDMINaDevice>, and
$<TARGET_FILE:MQT::CoreQDMI_DDSIM_Device> to use POST_BUILD and ensure the
target dependency is the ${TARGET_NAME} (or add add_dependencies(${TARGET_NAME}
<device-target>) if needed) so the device DLLs are copied deterministically
across generators.

---

Outside diff comments:
In `@python/mqt/core/plugins/qiskit/backend.py`:
- Around line 230-239: Normalize op_name to lowercase immediately after
retrieving it from op.name() and before any further checks or the alias lookup
so the case-insensitive mapping in _QDMI_TO_QISKIT_GATE_MAP works and deduping
doesn't produce different-cased duplicates; specifically, update the block that
sets op_name = op.name() and the subsequent lookup against
_QDMI_TO_QISKIT_GATE_MAP to use op_name = op.name().lower(), and apply the same
normalization to the later occurrence that currently performs a lookup/alias
remapping (the lines around where op_name is remapped again) so all comparisons
and remaps use the lowercased op_name consistently.
- Around line 63-64: The type annotations for the Qiskit↔QDMI mappings currently
only allow Instruction instances but the maps also store instruction classes
(e.g., MCXGate, MCPhaseGate); update the annotations to include both classes and
instances by importing typing.Type and typing.Union and changing occurrences
like dict[str, Instruction] or set[Instruction] to dict[str, Union[Instruction,
Type[Instruction]]] or set[Union[Instruction, Type[Instruction]]] (or use
Type[Instruction] where only classes are stored); apply this change to the
function return annotation that currently reads ") -> tuple[dict[str, set[str]],
dict[str, Instruction]]" and to the other map/type hints referenced around the
same module (the maps defined near lines ~86–87, ~158–159, and ~324–325) so all
places that can hold classes or instances accept Type[Instruction] as well as
Instruction.

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 `@test/python/plugins/qiskit/test_mock_backend.py`:
- Line 189: The zip in the dict comprehension that builds the outcome->count
mapping should use strict=True to ensure outcomes and counts_list have the same
length; update the expression that currently uses zip(outcomes, counts_list,
strict=False) to zip(outcomes, counts_list, strict=True) so any future length
mismatches are caught (look for the comprehension referencing outcomes and
counts_list in test_mock_backend.py).
- Around line 174-175: The code returns {} when self._num_clbits == 0 which
breaks the counts-sum invariant; change that branch to return {"": self._shots}
so the single empty-bitstring outcome accumulates all shots (use self._shots and
keep the empty-string key). Also update the zip call that iterates outcomes (the
one pairing sequences of length num_outcomes) to use zip(..., strict=True) to
enforce equal lengths and catch future mismatches; look for the zip(...) using
num_outcomes and replace with zip(..., strict=True).

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 `@docs/qdmi/driver.md`:
- Line 15: Fix the duplicated word in the qdmi Driver description: replace the
phrase "several devices preloaded devices" with "several preloaded devices" so
the sentence reads e.g. "MQT Core's QDMI Driver, qdmi::Driver, comes with
several preloaded devices that can be used directly." Ensure only the duplicated
word is removed and the class reference qdmi::Driver remains unchanged.

---

Outside diff comments:
In `@python/mqt/core/plugins/qiskit/backend.py`:
- Around line 61-100: The return and variable type hints must allow both
Instruction instances and Instruction classes: change
_build_gate_mappings_for_backend's return annotation to tuple[dict[str,
set[str]], dict[str, Instruction | type[Instruction]]], update the local
variable declaration operation_to_gate: dict[str, Instruction |
type[Instruction]], and rename the loop variable gate_instance to gate; also
update the module ClassVar _OPERATION_TO_GATE_MAP to ClassVar[dict[str,
Instruction | type[Instruction]]] and change _map_operation_to_gate's return
type to Instruction | type[Instruction] | None so all signatures reflect
heterogeneous values.

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 26: Replace the unhyphenated word "builtin" in the changelog entry "♻️
Build all builtin QDMI devices as shared libraries" with the correct compound
adjective "built-in"; edit the CHANGELOG entry text so it reads "♻️ Build all
built-in QDMI devices as shared libraries" to apply the hyphenation rule for
compound adjectives.