munich-quantum-toolkit · burgholzer · Sep 8, 2026 · Sep 8, 2026 · Sep 8, 2026 · Sep 8, 2026
diff --git a/.agent/audits/controlled-rotation-quality.csv b/.agent/audits/controlled-rotation-quality.csv
@@ -0,0 +1,145 @@
+axis,controls,angle,backend,synthesis_ms,raw_cx,raw_one_qubit,raw_depth,raw_cx_depth,optimized_cx,optimized_one_qubit,optimized_depth,optimized_cx_depth,max_operator_error
+rx,2,numeric,core,0.272258,4,6,10,4,4,5,9,4,2.1153413308660057e-16
+rx,2,numeric,qiskit,0.033504,8,15,18,8,8,7,13,8,3.298440587052649e-16
+rx,2,symbolic,core,0.365394,4,6,10,4,4,6,10,4,3.1401849173675503e-16
+rx,2,symbolic,qiskit,0.043104,8,15,18,8,8,9,15,8,3.3306690738754696e-16
+rx,3,numeric,core,0.382035,14,24,28,12,14,16,20,12,1.5331320705493272e-15
+rx,3,numeric,qiskit,0.069664,20,35,42,20,20,15,29,20,7.610516769633753e-16
+rx,3,symbolic,core,0.384211,14,24,28,12,14,20,24,12,1.2780539954372932e-15
+rx,3,symbolic,qiskit,0.086401,20,35,42,20,20,21,35,20,4.440892098500626e-16
+rx,4,numeric,core,0.433683,24,42,46,18,24,28,34,18,2.3361959382837703e-15
+rx,4,numeric,qiskit,0.189521,24,42,46,18,24,28,34,18,2.3361959382837703e-15
+rx,4,symbolic,core,0.448979,24,42,46,18,24,36,40,18,2.558937633260452e-15
+rx,4,symbolic,qiskit,0.087009,24,42,46,18,24,36,40,18,2.558937633260452e-15
+rx,5,numeric,core,0.26325,40,58,70,34,40,48,64,34,2.6801557172297703e-15
+rx,5,numeric,qiskit,0.074576,40,58,70,34,40,48,64,34,2.6801557172297703e-15
+rx,5,symbolic,core,0.265666,40,58,70,34,40,56,68,34,3.9563492527737815e-15
+rx,5,symbolic,qiskit,0.095857,40,58,70,34,40,56,68,34,3.9563492527737815e-15
+rx,6,numeric,core,0.350434,56,74,86,38,56,69,76,38,
+rx,6,numeric,qiskit,0.089969,56,74,86,38,56,69,76,38,
+rx,6,symbolic,core,0.342834,56,74,86,38,56,75,84,38,
+rx,6,symbolic,qiskit,0.101105,56,74,86,38,56,75,84,38,
+rx,7,numeric,core,0.426562,80,124,176,80,80,88,152,80,
+rx,7,numeric,qiskit,0.250914,80,124,176,80,80,88,152,80,
+rx,7,symbolic,core,0.934085,80,124,176,80,80,96,156,80,
+rx,7,symbolic,qiskit,0.306498,80,124,176,80,80,96,156,80,
+rx,8,numeric,core,1.081863,104,174,213,98,104,120,185,98,
+rx,8,numeric,qiskit,0.24325,104,174,200,89,104,120,170,89,
+rx,8,symbolic,core,1.085719,104,174,213,98,104,128,185,98,
+rx,8,symbolic,qiskit,0.315538,104,174,200,89,104,128,170,89,
+rx,9,numeric,core,1.210728,120,198,248,112,120,136,213,112,
+rx,9,numeric,qiskit,0.253954,120,198,248,108,120,136,209,108,
+rx,9,symbolic,core,1.215672,120,198,248,112,120,144,213,112,
+rx,9,symbolic,qiskit,0.306066,120,198,248,108,120,144,209,108,
+rx,10,numeric,core,1.361704,136,222,293,130,136,152,249,130,
+rx,10,numeric,qiskit,0.277426,136,222,280,121,136,152,234,121,
+rx,10,symbolic,core,1.386809,136,222,293,130,136,160,249,130,
+rx,10,symbolic,qiskit,0.329666,136,222,280,121,136,160,234,121,
+rx,16,numeric,core,2.290686,232,366,533,226,232,248,441,226,
+rx,16,numeric,qiskit,0.335074,232,366,520,217,232,248,426,217,
+rx,16,symbolic,core,2.301854,232,366,533,226,232,256,441,226,
+rx,16,symbolic,qiskit,0.378083,232,366,520,217,232,256,426,217,
+rx,32,numeric,core,4.60406,488,750,1173,482,488,504,953,482,
+rx,32,numeric,qiskit,0.472531,488,750,1160,473,488,504,938,473,
+rx,32,symbolic,core,4.589452,488,750,1173,482,488,512,953,482,
+rx,32,symbolic,qiskit,0.519923,488,750,1160,473,488,512,938,473,
+rx,64,numeric,core,9.109,1000,1518,2453,994,1000,1016,1977,994,
+rx,64,numeric,qiskit,0.707028,1000,1518,2440,985,1000,1016,1962,985,
+rx,64,symbolic,core,9.080984,1000,1518,2453,994,1000,1024,1977,994,
+rx,64,symbolic,qiskit,0.842309,1000,1518,2440,985,1000,1024,1962,985,
+ry,2,numeric,core,0.258562,4,4,8,4,4,4,8,4,2.220446049250313e-16
+ry,2,numeric,qiskit,0.036944,8,15,18,8,8,6,12,8,2.220446049250313e-16
+ry,2,symbolic,core,0.270545,4,4,8,4,4,4,8,4,2.220446049250313e-16
+ry,2,symbolic,qiskit,0.042288,8,15,18,8,8,6,12,8,3.3306690738754696e-16
+ry,3,numeric,core,0.350882,14,22,26,12,14,17,21,12,1.024360521199179e-15
+ry,3,numeric,qiskit,0.075584,20,35,42,20,20,14,28,20,2.220446049250313e-16
+ry,3,symbolic,core,0.355074,14,22,26,12,14,20,24,12,8.892273665708786e-16
+ry,3,symbolic,qiskit,0.095025,20,35,42,20,20,14,28,20,4.440892098500626e-16
+ry,4,numeric,core,0.406435,24,40,44,18,24,29,35,18,1.520397016336855e-15
+ry,4,numeric,qiskit,0.177777,24,40,44,18,24,29,35,18,1.520397016336855e-15
+ry,4,symbolic,core,0.408035,24,40,44,18,24,36,41,18,1.686470170725321e-15
+ry,4,symbolic,qiskit,0.207521,24,40,44,18,24,36,41,18,1.686470170725321e-15
+ry,5,numeric,core,0.562084,40,56,69,34,40,50,64,34,5.0035533349649525e-15
+ry,5,numeric,qiskit,0.197794,40,56,69,34,40,50,64,34,5.0035533349649525e-15
+ry,5,symbolic,core,0.558788,40,56,69,34,40,55,68,34,4.0974976367522364e-15
+ry,5,symbolic,qiskit,0.25357,40,56,69,34,40,55,68,34,4.0974976367522364e-15
+ry,6,numeric,core,0.719669,56,72,85,38,56,67,76,38,
+ry,6,numeric,qiskit,0.200482,56,72,85,38,56,67,76,38,
+ry,6,symbolic,core,0.345506,56,72,85,38,56,74,83,38,
+ry,6,symbolic,qiskit,0.095505,56,72,85,38,56,74,83,38,
+ry,7,numeric,core,0.440963,80,122,174,80,80,89,153,80,
+ry,7,numeric,qiskit,0.095392,80,122,174,80,80,89,153,80,
+ry,7,symbolic,core,0.42237,80,122,174,80,80,96,156,80,
+ry,7,symbolic,qiskit,0.108529,80,122,174,80,80,96,156,80,
+ry,8,numeric,core,0.506563,104,172,212,98,104,121,186,98,
+ry,8,numeric,qiskit,0.237873,104,172,199,89,104,121,171,89,
+ry,8,symbolic,core,1.086358,104,172,212,98,104,128,186,98,
+ry,8,symbolic,qiskit,0.274962,104,172,199,89,104,128,171,89,
+ry,9,numeric,core,1.203495,120,196,247,112,120,137,214,112,
+ry,9,numeric,qiskit,0.243874,120,196,247,108,120,137,210,108,
+ry,9,symbolic,core,1.210119,120,196,247,112,120,144,214,112,
+ry,9,symbolic,qiskit,0.301426,120,196,247,108,120,144,210,108,
+ry,10,numeric,core,1.335944,136,220,292,130,136,153,250,130,
+ry,10,numeric,qiskit,0.25957,136,220,279,121,136,153,235,121,
+ry,10,symbolic,core,1.356264,136,220,292,130,136,160,250,130,
+ry,10,symbolic,qiskit,0.319554,136,220,279,121,136,160,235,121,
+ry,16,numeric,core,2.213934,232,364,532,226,232,249,442,226,
+ry,16,numeric,qiskit,0.346178,232,364,519,217,232,249,427,217,
+ry,16,symbolic,core,2.235966,232,364,532,226,232,256,442,226,
+ry,16,symbolic,qiskit,0.369299,232,364,519,217,232,256,427,217,
+ry,32,numeric,core,2.109981,488,748,1172,482,488,505,954,482,
+ry,32,numeric,qiskit,0.205905,488,748,1159,473,488,505,939,473,
+ry,32,symbolic,core,2.101613,488,748,1172,482,488,512,954,482,
+ry,32,symbolic,qiskit,0.233938,488,748,1159,473,488,512,939,473,
+ry,64,numeric,core,4.250746,1000,1516,2452,994,1000,1017,1978,994,
+ry,64,numeric,qiskit,0.349938,1000,1516,2439,985,1000,1017,1963,985,
+ry,64,symbolic,core,4.418747,1000,1516,2452,994,1000,1024,1978,994,
+ry,64,symbolic,qiskit,0.359011,1000,1516,2439,985,1000,1024,1963,985,
+rz,2,numeric,core,0.094,4,4,8,4,4,4,8,4,0.0
+rz,2,numeric,qiskit,0.065872,4,4,8,4,4,4,8,4,0.0
+rz,2,symbolic,core,0.107504,4,4,8,4,4,4,8,4,0.0
+rz,2,symbolic,qiskit,0.074577,4,4,8,4,4,4,8,4,0.0
+rz,3,numeric,core,0.146417,14,22,26,12,14,17,21,12,9.71445146547012e-16
+rz,3,numeric,qiskit,0.070896,14,22,26,12,14,17,21,12,9.71445146547012e-16
+rz,3,symbolic,core,0.146593,14,22,26,12,14,20,24,12,8.473409486550037e-16
+rz,3,symbolic,qiskit,0.081153,14,22,26,12,14,20,24,12,8.473409486550037e-16
+rz,4,numeric,core,0.183121,24,40,44,18,24,29,35,18,1.3092278833360677e-15
+rz,4,numeric,qiskit,0.075681,24,40,44,18,24,29,35,18,1.3092278833360677e-15
+rz,4,symbolic,core,0.438195,24,40,44,18,24,36,41,18,1.6136471996107273e-15
+rz,4,symbolic,qiskit,0.213825,24,40,44,18,24,36,41,18,1.6136471996107273e-15
+rz,5,numeric,core,0.553923,40,56,69,34,40,48,64,34,2.538614803067886e-15
+rz,5,numeric,qiskit,0.199377,40,56,69,34,40,48,64,34,2.538614803067886e-15
+rz,5,symbolic,core,0.270114,40,56,69,34,40,55,68,34,2.871672278914067e-15
+rz,5,symbolic,qiskit,0.099201,40,56,69,34,40,55,68,34,2.871672278914067e-15
+rz,6,numeric,core,0.355427,56,72,85,38,56,69,76,38,
+rz,6,numeric,qiskit,0.095633,56,72,85,38,56,69,76,38,
+rz,6,symbolic,core,0.36133,56,72,85,38,56,74,83,38,
+rz,6,symbolic,qiskit,0.102721,56,72,85,38,56,74,83,38,
+rz,7,numeric,core,0.447603,80,122,174,80,80,89,153,80,
+rz,7,numeric,qiskit,0.101792,80,122,174,80,80,89,153,80,
+rz,7,symbolic,core,0.453491,80,122,174,80,80,96,156,80,
+rz,7,symbolic,qiskit,0.112609,80,122,174,80,80,96,156,80,
+rz,8,numeric,core,0.528067,104,172,212,98,104,121,186,98,
+rz,8,numeric,qiskit,0.108993,104,172,199,89,104,121,171,89,
+rz,8,symbolic,core,0.53146,104,172,212,98,104,128,186,98,
+rz,8,symbolic,qiskit,0.119873,104,172,199,89,104,128,171,89,
+rz,9,numeric,core,0.60346,120,196,247,112,120,137,214,112,
+rz,9,numeric,qiskit,0.275169,120,196,247,108,120,137,210,108,
+rz,9,symbolic,core,1.363753,120,196,247,112,120,144,214,112,
+rz,9,symbolic,qiskit,0.341538,120,196,247,108,120,144,210,108,
+rz,10,numeric,core,1.525946,136,220,292,130,136,153,250,130,
+rz,10,numeric,qiskit,0.288994,136,220,279,121,136,153,235,121,
+rz,10,symbolic,core,1.479721,136,220,292,130,136,160,250,130,
+rz,10,symbolic,qiskit,0.145633,136,220,279,121,136,160,235,121,
+rz,16,numeric,core,1.075383,232,364,532,226,232,249,442,226,
+rz,16,numeric,qiskit,0.145457,232,364,519,217,232,249,427,217,
+rz,16,symbolic,core,1.07559,232,364,532,226,232,256,442,226,
+rz,16,symbolic,qiskit,0.163921,232,364,519,217,232,256,427,217,
+rz,32,numeric,core,2.147598,488,748,1172,482,488,505,954,482,
+rz,32,numeric,qiskit,0.482483,488,748,1159,473,488,505,939,473,
+rz,32,symbolic,core,4.725454,488,748,1172,482,488,512,954,482,
+rz,32,symbolic,qiskit,0.505475,488,748,1159,473,488,512,939,473,
+rz,64,numeric,core,9.260025,1000,1516,2452,994,1000,1017,1978,994,
+rz,64,numeric,qiskit,0.746389,1000,1516,2439,985,1000,1017,1963,985,
+rz,64,symbolic,core,9.402778,1000,1516,2452,994,1000,1024,1978,994,
+rz,64,symbolic,qiskit,0.842181,1000,1516,2439,985,1000,1024,1963,985,
diff --git a/.agent/audits/controlled-synthesis-test-runtime.md b/.agent/audits/controlled-synthesis-test-runtime.md
@@ -0,0 +1,49 @@
+# Controlled synthesis test runtime
+
+Status: implemented and validated. Baseline:
+`1f25f40667079f909685992653549b12ce0e0be6` with the pending MCY support, shared
+Pauli tests, and rotation CX budgets. The open PR is
+[#2467](https://github.com/munich-quantum-toolkit/core/pull/2467).
+
+## Scope and contract
+
+Speed up `test_multi_controlled_decomposition.cpp` without removing control
+widths, numeric or runtime angles, gate-count bounds, exact-phase comparisons,
+borrowed-control restoration, or input/output IR and linearity verification.
+Production synthesis and the Python test matrix are unchanged.
+
+## Findings
+
+The single-pass helper verified its output twice: once in `PassManager::run` and
+again in an explicit `verify` call. Sampling the rotation resource test
+attributed 658 samples to the first check and 677 to the second, compared with
+712 in greedy rewriting. The helper now explicitly enables the pass manager's
+verifier and retains the separate output linearity check. Input verification is
+unchanged. Both output checks covered the same module after the same pass.
+
+The large Pauli full-operator tests spend most of their time in DD matrix
+multiplication. Reusing packages solely to avoid allocations is not justified by
+this profile. Switching the three matrix-only helpers to the existing
+unitary-simulation DD configuration showed no benefit: the three-run median for
+the 25 Pauli-at-eight-controls and numeric/runtime rotation tests changed from
+5.276 to 5.294 seconds. This experiment was reverted.
+
+## Validation
+
+The baseline Debug binary passed all 297 decomposition tests in 42.4 seconds on
+macOS ARM64 with AppleClang 21 and LLVM/MLIR 23.1. The run included a short
+sampling interval, so it is diagnostic rather than a controlled speed ratio.
+
+A subsequent unprofiled three-run comparison of
+`MultiControlledDecompositionTest.RotationsUseLinearResourcesWithoutExtraQubits`
+used the original binary and the rebuilt single-verifier binary serially, with
+no concurrent build. The median fell from 2.035 to 1.477 seconds (27%). All six
+runs passed with all 96 axis, width, and angle-kind combinations retained. The
+34-test MCY and rotation-resource filter also passed three times per binary. Its
+median fell from 10.528 to 8.910 seconds (15%). Compilation and profiling time
+are excluded from test durations.
+
+The final rebuilt binary passes all 297 tests. The full changed-file
+`uvx nox -s cpp-lint -- ec799daa09f855bd0edcbc5592a5fedd90836516` check reports
+zero findings, and `uvx nox -s lint` passes. The final full-suite run overlapped
+with C++ lint and is not used for a whole-suite speedup claim.
diff --git a/.agent/audits/pr2467-controlled-synthesis-review.md b/.agent/audits/pr2467-controlled-synthesis-review.md
@@ -0,0 +1,150 @@
+# Controlled-synthesis audit and quality comparison
+
+Status: applied. Date: 2026-09-08. Original baseline:
+`00a214ec3f49aab0eeee0352a05ef817100e8c55`. Changes rebased onto main
+`4c5e45855e5bb50c42b68ddbf4f9a4dababb8737`.
+
+## Findings and disposition
+
+- **Identity modifier evaluation:** wide X/Y/Z synthesis can leave empty
+  `qco.ctrl` bodies after angle folding. The original head's DD consumer
+  rejected these valid identity regions. Main's shared `composeBodyMatrix` fix
+  in #2464 resolves this on rebase. No duplicate or Y-specific workaround is
+  needed. The 63-control coherent-state regression now passes without
+  canonicalization.
+- **Unreachable HP24 machinery:** `mczCoreForWidth` uses specialized synthesis
+  below five controls and SP22 through 32. HP24 now asserts its actual minimum
+  of 33 controls and directly selects one dirty helper for odd widths, two for
+  even widths. Removed the inactive small-width table, ripple incrementer,
+  recursive relative-phase planner, thread-local cache, and estimate branches.
+  Active half-MCX widths exceed 11 and incrementer widths exceed 10, so those
+  removed alternatives had no production caller. Changing the crossover must
+  revisit this limit. This is a maintenance improvement, not a speed claim.
+- **Wide-state oracle and coverage:** retained existing SP22 samples and added
+  32/33/34, 47/48, and 63/64 controls for X/Y/Z. Compare the phase-sensitive
+  norm of `actual - expected` at `1e-11`, rather than DD node identity. The
+  coherent helper scopes DD arithmetic tolerance to `1e-15` and restores it
+  afterward; the default merging tolerance accumulated about `1e-9` error at 32
+  controls. These selected states are not a full-operator bound. Independent
+  numeric and runtime rotation matrices now cover 2 through 10 controls.
+- **Duplicate remapping:** remap each generated rotation half-plan in place;
+  remove unused `GateEmitter` remapping. This removes a second plan allocation
+  and move loop. Plans remain the sole owner of wire remapping.
+- **Stale CLI descriptions:** both compiler options now name Y and rotations;
+  the generated pass description names the active HP24 dirty-helper choice.
+
+Source:
+`mlir/lib/Dialect/QCO/Transforms/Decomposition/DecomposeMultiControlled.cpp`.
+Regression tests: the corresponding
+`mlir/unittests/Dialect/QCO/Transforms/Decomposition/test_multi_controlled_decomposition.cpp`.
+
+## Contracts retained
+
+The modifier verifier owns restrictions on classical support operations. No
+second dependency walker is needed. Balanced halves borrow opposite controls and
+restore them coherently. Arbitrary relative-phase MCX replacements have not been
+proved safe in the four-MCX rotation shell. RX uses Hadamard conjugation of RZ;
+Y uses target S conjugation of X. Controlled `2*pi` rotations retain their
+conditional minus sign. Native-target and minimum-width policies still apply. No
+new cache, public API, or dependency is justified.
+
+## Reproducible Qiskit comparison
+
+Run `uv run --no-sync python test/bench/compare_controlled_rotations.py` with
+the locally built package and Qiskit 2.5.2. The script records numeric and
+symbolic RX/RY/RZ at 2, 3, 4, 5, 6, 7, 8, 9, 10, 16, 32, and 64 controls. Raw
+rows from this run are in
+[controlled-rotation-quality.csv](controlled-rotation-quality.csv). Core was
+built with Clang/LLVM 23.1, Release, ThinLTO, and mold on ARM64.
+
+Both methods use exactly the input qubits. Nine-sample median synthesis times
+exclude one warmup, input creation, import/export, basis conversion, and
+routing. Both circuits receive the same `u,cx` normalization, then level-3
+optimization with seed zero and `qubits_initially_zero=False`. The CSV includes
+CX count, one-qubit count, total depth, CX depth, and synthesis time. Numeric
+and symbolic operators through five controls pass phase-sensitive checks,
+including symbolic binding at `2*pi`. Timings are local sequential measurements,
+not an end-to-end compilation comparison or stable performance guarantee.
+
+Numeric post-optimization examples (Core / Qiskit):
+
+| Gate | Controls |          CX |       Depth |  Synthesis ms |
+| ---- | -------: | ----------: | ----------: | ------------: |
+| RX   |        2 |       4 / 8 |      9 / 13 | 0.272 / 0.034 |
+| RY   |        3 |     14 / 20 |     21 / 28 | 0.351 / 0.076 |
+| RY   |        8 |   104 / 104 |   186 / 171 | 0.507 / 0.238 |
+| RY   |       16 |   232 / 232 |   442 / 427 | 2.214 / 0.346 |
+| RY   |       64 | 1000 / 1000 | 1978 / 1963 | 4.251 / 0.350 |
+
+Core saves CX gates for RX/RY at two and three controls. Other sampled CX counts
+match, including all RZ cases and symbolic angles. Core's larger circuits are up
+to 15 layers deeper. Core synthesis is slower in this measurement. Earlier
+claims of uniformly faster Core synthesis included basis lowering in Qiskit's
+timed work and are superseded by this comparison.
+
+## Improvement beyond Qiskit: measured candidate, deferred implementation
+
+Helper order affects scheduling even when CX count stays fixed. In an isolated
+prototype, reverse the selected dirty-helper wires in both balanced half-MCXs,
+keeping controls in order. Use Qiskit's exact `synth_mcx_n_dirty_i15` to
+construct those halves, and the same four quarter-angle rotations and
+normalization as above. This produces these numeric RY depths:
+
+| Controls | Core | Qiskit | Reversed-helper prototype | CX (all three) |
+| -------- | ---: | -----: | ------------------------: | -------------: |
+| 9        |  214 |    210 |                       204 |            120 |
+| 10       |  250 |    235 |                       232 |            136 |
+| 16       |  442 |    427 |                       388 |            232 |
+| 32       |  954 |    939 |                       804 |            488 |
+| 64       | 1978 |   1963 |                      1636 |           1000 |
+
+At eight controls, balanced reversed helpers give depth 180, worse than Qiskit's
+171. A 5+3 split instead reduces CX from 104 to 96 but has depth 185. Neither
+candidate dominates at every width. The reversed balanced prototype at nine
+controls and both eight-control splits pass full phase-sensitive operator
+comparisons with maximum element error below `1.2e-14`.
+
+Minimal reproduction, after importing `QuantumCircuit`, `transpile`, and
+`qiskit.synthesis.synth_mcx_n_dirty_i15`:
+
+```python
+k = 64
+first = (k + 1) // 2
+halves = []
+for start, count in ((0, first), (first, k - first)):
+    plan = synth_mcx_n_dirty_i15(count)
+    spare = [i for i in range(k) if i < start or i >= start + count]
+    spare = spare[: plan.num_qubits - count - 1]
+    halves.append((plan, list(range(start, start + count)) + [k] + spare[::-1]))
+circuit = QuantumCircuit(k + 1)
+for _ in range(2):
+    circuit.compose(*halves[0], inplace=True)
+    circuit.ry(-0.73 / 4, k)
+    circuit.compose(*halves[1], inplace=True)
+    circuit.ry(0.73 / 4, k)
+output = transpile(
+    circuit,
+    basis_gates=["u", "cx"],
+    optimization_level=3,
+    seed_transpiler=0,
+    qubits_initially_zero=False,
+)
+assert output.count_ops()["cx"] == 1000
+assert output.depth() == 1636
+```
+
+This establishes room beyond Qiskit's current public synthesis output, not
+optimality or a ready Core patch. Before implementation, measure Core's own
+helper ordering across all axes, symbolic expressions, and routing targets;
+retain the exact-restoration tests. A width-specific split policy needs an
+explicit objective because CX count and depth disagree at eight controls. The
+existing construction has linear CX count; results requiring additional
+clean/dirty qubits do not establish an improvement under this no-extra-qubit
+contract.
+
+## Validation
+
+The native decomposition binary passes all 303 tests, including both active HP24
+dirty-helper modes and numeric/runtime full operators through ten controls. The
+benchmark completes all 144 backend rows. Final lint and Python validation are
+recorded in the implementation plan.