[cjdoris]

Redesign pyconvert around a global Python source-type DAG, destination scopes, stable rule IDs, cached selections, and explicit conversion contexts. Canonical conversions prefer native materialized containers while generic iterators remain lazy. Add source-specific preferences, lazy public type aliases and shorthand such as "list" => PyList and "None" => Missing.

Rules and public pytryconvert share a normalized result convention: nothing means failure, ordinary successes are raw values, and successful nothing/Some values are wrapped once. After checking for failure, something(result) recovers the value. This avoids unnecessary result allocations while preserving concrete and union return bounds. Exceptions propagate.

Restrict generic floating-point coercions to Python float/complex sources so integer overflow cannot succeed after rounding through a float. Integral/rational sources retain exact intermediates; supported NumPy scalars retain dedicated rules. Update conversion documentation, migration guidance, and invariant tests.

Validation: 1,812 focused conversion/wrapper/Julia-wrapper checks passed with NumPy installed (one existing optional-dependency skip); 1,694 passed with bare system Python (four expected optional-dependency skips). Full-suite and supported-version coverage run in CI. NumPy-specific tests respect its optional status in external-interpreter jobs.

Reproducible benchmark scripts are included; generated outputs were removed. The latest direct allocation check saves 16 bytes/one allocation for public String, BigInt and array try-conversion; concrete Int/Float64 remain allocation-free. Earlier original-v1 comparisons showed inferred large-vector conversion improving by roughly one third, and array allocation totals matching original v1 after minimal rule-result wrapping. The remaining major measured regression is cold registry rebuilding (~2.6 ms versus ~28 μs in the synthetic fixture), plus slower simple selection misses. Warm calls do not rebuild the DAG.

Related to #661.