Use native NNS distance by default package-wide

[OVVO-Financial]

Motivation

• Align the Python package with the native blended NNS default so that any public API or internal delegation that omits dist uses the native NNS formula (sum(abs(z) + z**2)) rather than implicitly retaining historical L2 behavior.
• Ensure parity helpers and R bridge code represent None/NULL explicitly so cache keys and R calls distinguish the native default from intentionally-requested L2.
• Add invariant tests to prevent accidental reintroduction of hard-coded dist="L2" delegations in ordinary package source.

Description

• Switched internal delegations that previously forced dist="L2" to use dist=None (or omit the dist kwarg) in derivative and VAR call paths, primarily in src/nns/diff.py and src/nns/var.py.
• Updated R parity helpers in tests/_r.py to accept dist: str | None = None, include "dist": dist in the helper argument dictionaries, decode args$dist to dist_arg (NULL vs character) inside embedded R scripts, and pass dist = dist_arg by name into NNS.reg, NNS.stack, and NNS.boost calls.
• Rewrote general parity tests under tests/parity/ so default parity comparisons use Python dist=None and R dist = NULL instead of implicit L2; preserved and kept dedicated explicit-L2 tests unchanged where they assert explicit L2 behavior.
• Added focused invariant tests in tests/invariants/test_native_distance_defaults.py that (1) assert derivative helpers delegate with native None (monkeypatched checks) and (2) scan package source AST to fail if an ordinary package call delegates with a hard-coded dist='L2'.
• Kept validator/implementation locations and dedicated explicit-L2 tests unchanged; docstrings and R manual pages that describe explicit dist="L2" remain as documentation.

Testing

• ruff check . completed successfully.
• python -m compileall -q src tests scripts completed successfully (AST/compile sanity OK).
• Ran unit/test entrypoints; many pytest runs could not execute here because required runtime dependencies (notably numpy) and R (Rscript) are not available in this environment, so live R parity and full pytest runs are blocked.
• python scripts/regenerate_r_cache.py --fresh -- tests/parity could not run because Rscript is not on PATH in this environment and so the R cache could not be regenerated here; the PR includes the helper changes required prior to regeneration but the actual cache rebuild must be performed in the authorized CI/regeneration environment and committed afterwards.
• mypy and python -m build were attempted but were blocked or produced missing-import diagnostics due to the same unavailable build/dependency tooling; these checks should be re-run in CI where dependencies are installed.

Notes on remaining hard-coded dist="L2" occurrences: the only allowed remaining matches are explicit L2 tests and documentation (for example tests/invariants/test_native_distance_defaults.py explicit-L2 assertions and R manual pages); there are no remaining hard-coded dist="L2" delegations inside ordinary package source after this change, and the new invariant test will detect regressions.

If you want, I can (1) run the official R cache regeneration in the proper environment and update tests/_r_cache.json, or (2) open follow-up CI job output to validate live R parity and cache-only parity once the pipeline runs in CI with numpy and R installed.

---

Codex Task