66922b46

By: ValentinHirschi <valentin.hirschi@gmail.com>

differentiate active sampling coordinates with fixed prepared parameters

Select ordered active columns in the existing eager/dual Jacobian evaluator.
Pass statically zero identity seeds to Symbolica so unused prepared-only
singular derivatives cannot contaminate valid map derivatives. Keep singular
active derivatives as typed numerical failures and preserve physical evaluator
behavior with empty masks. No second evaluator/cache or stored layout change.

Validation: 71 native/eager/map/partition and physical dual-shape checks passed;
saved-state acceptance passed (loaded amplitude and conditional-cut maps,
ordinary integration and exact resume). Core/API Python-feature cargo checks,
formatting, diff check and all-target clippy passed with no warnings in touched
files; unrelated pre-existing warnings remain. gammaloop_state layout unchanged.
The joint map and compact-support handling are still pending.

User prompt:
Just a status question—please continue your current orchestration unchanged: where does the joint two-normal density sit in the implementation sequence, and what prerequisites remain before you reach it? (direct-H focusing alone is not expected to bound the previously identified corner’s weights.)