8a28de51

By: lcnbr <im@lcnbr.ch>

Reduce build dependencies and evaluator generation/runtime costs

Prune unused dependency branches and consolidate capability features while
preserving numerical backends and state formats. Use the published Python
stub derive crate and its signature macros, pin Momtrop fixes upstream, and
disable its optional DOT reader.

Select reachable aliases once per orientation, remap literal handles by
lookup, and collect weighted selector contributions into lazy shared regions.
Keep alias tags and numerator factorization intact.

Build summed function maps from selected residue-map entries so constant
selectors simplify before linearization. Bind undirected evaluator signs
to zero while preserving partial graph-selection semantics.

Scope foreign float metadata before exporting Python stubs, refresh reviewed
architecture scopes, and apply both existing performance tolerance floors.
Run feature-enabled exporter regression tests in documentation CI.
Honor complex scientific-format precision while retaining strict numeric
reference checks when evaluator ordering changes final floating-point bits.

Extend the paired CI performance gate with complementary alias-heavy and
weighted-guard fixtures. Build and benchmark every evaluator method, gate
construction and runtime separately, and verify their shared physical results.
Keep fresh generation, memory budgets and the four-minute confirmation cap.

Repair main CI tool prerequisites and macOS watcher/path fixtures. Use
authenticated prebuilt Jujutsu, preserve watcher feature isolation, and replace
a one-element Vakint loop with its identical guarded update for new Clippy.
Reuse authenticated NixCI artifacts with bounded build concurrency, store
collection and output roots retained through publication.

Co-authored-by: Kaapo Seppänen <seppanenkaapo@gmail.com>