Release wave: QUBODrivers 0.6 compat + typed attributes + benchmark conformance

[bernalde]

Part of the benchmarking-readiness epic: JuliaQUBO/QUBO.jl#40 (WS2 — coordinated driver release wave). Blocked by the QUBODrivers metadata/seed standardization and conformance-suite issues on that epic (WS1) — adopt both in a single PR here so the package is only touched once.

Problem

QUBO.jl v0.6 requires QUBODrivers = "0.6" / QUBOTools = "0.13". CIMOptimizer.jl's current compat is:

QUBODrivers = "0.5"
(QUBOTools: via QUBODrivers)

The intersection with QUBO.jl v0.6 (and with the QUBOLib-served QOBLib benchmarking stack) is empty — this package cannot be installed alongside the rest of the ecosystem. A scheduled ecosystem-resolution canary is being added to QUBO.jl that will stay red until this lands.

Package-specific work

• All ~50 hyperparameters go through untyped RawOptimizerAttribute keys — fine for the long tail, but promote the campaign-relevant ones (num_runs, time limit, seed if the cim-optimizer backend supports one) to typed @setup attributes.
• Timing comes from solver.result["time"] — verify what it measures (PyTorch solve loop vs including tensor setup) and map it to "effective" accordingly; document under the namespaced breakdown if setup is significant.
• PyTorch dependency makes this the heaviest Python environment in the wave: confirm CondaPkg resolution alongside PySA/DWave in one environment, or document that it needs an isolated environment (the harness supports per-solver environments either way, but it must be known).

Common work (all drivers in the wave)

1. Bump compat: QUBODrivers = "0.6" (drop pre-0.5 entries unless there's a reason to keep them) and QUBOTools = "0.13" where it is a direct dependency. Fix any breakage from the upgrade.
2. Adopt the QUBODrivers metadata schema: metadata["time"]["effective"] (backend solve time) always set; seeds recorded under metadata["seeds"]; backend name/version under metadata["backend"].
3. Declare capability traits honestly (supports_seed, honors_final_reads, enforces_time_limit).
4. Adopt the standard RandomSeed attribute where the backend supports seeding (alias any existing seed attribute to it; keep the old one as deprecated).
5. CI runs the upgraded QUBODrivers.test(Optimizer) with benchmark_conformance = true.
6. Tag a release and verify it registers (General registry compat must match).

Checkpoints (anti-drift)

• Before starting: wait for the reference driver run posted on the epic (first wave driver, suggested MQLib.jl) and use its conformance output as the template — deviations should be deliberate, not accidental.
• Mid-PR: post the conformance-suite output here (all sets, including which are trait-skipped and why).
• Before merge: resolve a temp environment with QUBO@0.6 + this branch via Pkg.develop — the actual coexistence check, not just compat-file edits.

Acceptance criteria

• Registered release with QUBODrivers 0.6 / QUBOTools 0.13 compat, conformance suite green in CI, package resolves alongside QUBO.jl v0.6.

[bernalde]

Reference driver checkpoint is ready enough to start this package's wave PR.

MQLib status:

• Add QUBODrivers 0.6 benchmark conformance MQLib.jl#27 merged into main.
• Reference checkpoint posted on the epic: Epic: Benchmark-ready JuliaQUBO ecosystem for the QOBLib campaign QUBO.jl#40.
• MQLib release-prep PR Prepare v0.7.0 release MQLib.jl#28 is open, CI-green, and bumps the release to v0.7.0; registration is still pending.
• Final reference stack uses QUBODrivers = "0.6.1" and QUBOTools = "0.13".
• Reference conformance result: supports_seed=true, honors_final_reads=true, enforces_time_limit=true, Benchmark Conformance 14/14, no trait-skipped benchmark sets.

Implementation guidance for CIMOptimizer.jl:

• Target QUBODrivers = "0.6.1"; add QUBOTools = "0.13" only if it is a direct dependency.
• Promote campaign-relevant raw attributes to typed setup attributes, especially run count, time limit, and seed if the backend supports deterministic seeding.
• If the backend cannot provide a reproducible seed path, declare supports_seed = false rather than weakening the conformance checks.
• Verify what solver.result["time"] measures and map backend solve time to metadata["time"]["effective"]; put setup or tensor-transfer timing in a namespaced breakdown if needed.
• Check whether the PyTorch/Python environment can coexist with DWave and PySA; document isolated-environment needs if coexistence is not practical.

Use MQLib.jl#27/#28 as the reference shape: compat bump, metadata schema, trait declarations, unconditional QUBODrivers.test(...; benchmark_conformance = true), conformance transcript, QUBO v0.6 coexistence check, then release/registration.

[bernalde]

Completed by #19 and released as v0.2.1. General registry now includes CIMOptimizer v0.2.1 with QUBODrivers 0.6.1 compat.