[burgholzer]

🤖 AI text below 🤖

Description

Replace nlohmann/json with stock LLVM JSON parsing, values, and streams across benchmarks, DD statistics, QDMI configuration and bindings, and IQM export. No custom JSON support library or compatibility parser remains.

Native benchmark factories, raw-count evaluation, JSON parsing, and Shor execution return owned LLVM errors. Native probability and QAOA cutSize return scalars with a binary-outcome/width precondition; Python validates those inputs. Evaluation checks every count key, including zero-weight keys, and rejects empty, zero-total, or overflowing counts.

Evaluation results own their validated shots total in C++ and Python. JSON formatting is private and infallible after manifest verification and successful evaluation. The public evaluationToJSON(caseId, shots, result) API is removed; callers use evaluateJSON(manifest, counts). ModularMultiplier reuses the result total. Native Shor execution borrows a valid, nonempty, nonthrowing callback through llvm::function_ref; Python callback exceptions retain their identity.

Bindings use nanobind for error handling, hashing, and warnings. Upstream's direct typed benchmark-generation path is preserved. CoreBench follows LLVM's EH/RTTI policy. The QDMI driver copies and SC tests match the SDK's RTTI setting. Benchmark bindings retain nanobind's RTTI and consume errors through LLVM's out-of-line APIs, so unoptimized modules also import correctly. Static CoreBench linkage preserves LLVM error-type identity; the MSVC project-options guard prevents /EHsc from overriding LLVM's no-exceptions policy. Existing QDMI exception boundaries remain. The C++ lint wrapper checks cpp-linter's finding count because its local backend can exit zero with findings.

Benchmark/configuration JSON uses LC_NUMERIC=C; Core does not change process locale. DD statistics and IQM numeric output use locale-independent APFloat. Allocation failure is outside the recoverable-error contract. The broader compiler/DD/device exception migration remains separate.

Base: 1d384a5d23aa1f02b26fd5684dab586ef26167fe. Native validation revision: ef4d19c2e35ed14118db7bc61a58e9e1acd61f8d. Guidance-only head: bf004898a2f489d22bea66a48fbf3a14aa2ef5f6.

Claim	Evidence
Input and numerical contracts remain checked	4,096 native tests pass in both GCC Debug and Clang ThinLTO Release, and all 2,042 Python tests pass. The JSON regression uses independently calculated TVD/fidelity.
Unoptimized LLVM consumers link and import	The prior head fails locally with the same missing LLVM typeinfo as the three failed hosted jobs. The corrected Debug native build passes; the Debug benchmark module imports and passes 85 Python tests. Malformed JSON keeps its ValueError and source diagnostic; overflow and callback-identity checks pass. All 183 focused QDMI coverage tests pass.
Shot totals belong to evaluation	Six targeted checks fail when propagation is replaced with zero, covering generic, QAOA, Shor, and JSON results; restored code passes. Generated stubs expose all three shots properties.
Native callbacks and evaluation remain exception-free	All three CoreBench compile units use no-EH/no-RTTI flags. Emitted LLVM IR has no invoke/landingpad instructions. Callback-error identity tests pass.
Keep ordered counts	Three alternating public-Python comparisons found unordered_map plus deterministic traversal about 15% slower for QFT4096 and 32% slower for QAOA4096; Shor improved 7%. Native phase probes isolate the added sort cost. The prototype is discarded.
Memory measures are distinct	For QFT4096, map/hash allocation traffic is 447/490 KB, peak tracked live memory 447/422 KB, retained memory zero, and process peak RSS including profiling about 34 MB. No whole-program speedup is claimed for the retained changes.
Final local checks pass	Full changed-file C++ lint, repository lint, stub generation, and specialist Ponytail review pass. The new commit is signed and verified. Stubs are unchanged. AGENTS.md asks for configurations relevant to the change and local reproduction before deleting build settings. It does not require both Debug and Release or CI monitoring. Repository lint passes after the guidance edit.

Local execution was on Linux. No new platforms, supported input shapes, adaptive execution, pulse functionality, validated-input framework, or custom container/caster is introduced.

GPT-6 via Codex assisted implementation, validation, review, and this description. Human review remains required before acceptance.

Checklist

• The pull request only contains commits that are focused and relevant to this change.
• I have added appropriate tests that cover the new/changed functionality.
• I have updated the documentation to reflect these changes.
• The changes follow the project's style guidelines and introduce no new warnings.
• The changes are fully tested and pass the CI checks.
• I have reviewed my own code changes.

If PR contains AI-assisted content:

• Any agent that created, edited, or submitted GitHub content was explicitly authorized for that scope, as required by our AI Usage Guidelines.
• Every agent-authored or agent-edited public text body begins with the visible disclosure 🤖 *AI text below* 🤖 (titles are exempt).
• I have disclosed AI assistance in the PR description.
• I confirm that I have personally reviewed and understood all AI-generated content, and accept full responsibility for it.

[codecov]

Codecov Report

❌ Patch coverage is 94.83827% with 75 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
mlir/bench/MQTCoreBench.cpp	66.9%	35 Missing ⚠️
src/bench/JSON.cpp	96.3%	22 Missing ⚠️
src/qdmi/devices/sc/Configuration.cpp	91.7%	14 Missing ⚠️
src/bench/QFTAdder.cpp	88.8%	3 Missing ⚠️
src/qdmi/driver/DeviceRegistry.cpp	98.0%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!

[burgholzer]

@denialhaag @simon1hofmann FYI: this PR does two things

• it migrates MQT Core away from nlohmann_json towards the builtin LLVM JSON capabilities. One dependency less 🎉
• it takes a first stab at the refactoring of the benchmark library to make it exception free. I hope this makes the migration in #2545 slightly smaller.

I tried to push Codex quite a bit to redesign/refactor the APIs to be more efficient and clean. 🧽

[simon1hofmann]

@denialhaag @simon1hofmann FYI: this PR does two things

• it migrates MQT Core away from nlohmann_json towards the builtin LLVM JSON capabilities. One dependency less 🎉
• it takes a first stab at the refactoring of the benchmark library to make it exception free. I hope this makes the migration in ♻️ Consolidate native Python bindings #2545 slightly smaller.

I tried to push Codex quite a bit to redesign/refactor the APIs to be more efficient and clean. 🧽

Pretty nice 😎