v1.106.0

Fixed

• Bench.optimize() never passed const_vars to the worker: _run_optuna_job folded the constants into the cache key but submitted only the trial-suggested values as job_args, so every Optuna trial silently ran with the worker class's parameter defaults for all const_vars. Constants are now merged into the submitted job_args (mirroring the sweep path's WorkerJob.setup_hashes); trial-suggested values keep precedence since _resolve_optimize_vars already strips colliding const entries. Regression coverage in test/test_optimize.py::TestConstVars for both the plain and aggregate/repeats>1 branches.
• CACHE_VERSION bumped to "4": because the old cache key already included the constants, any cached optimize() entries produced with non-default const_vars hold values actually computed with worker defaults — wrong data under a correct-looking key, indistinguishable on disk from good entries. The bump wipes the cache tree on first use of the new version so the fixed code can never warm-start from poisoned entries.