perf: select profitable compiled distance views

[isty2e]

Summary

• record whether a compiled geometry plan can amortize candidate-structure traversal
• use encoded distance views for array, tuple, and record spaces while keeping scalar and permutation spaces on their lower-overhead validated paths
• lock the dispatch boundary with focused tests and update the changelog

Why

Encoded snapshots avoid repeated composite traversal, but they add projection and alignment work when the existing geometry is already direct. In a 30-repeat rotating comparison, the selected path reduced the continuous workload from 129.50 to 113.88 us/eval relative to forcing scalar encoding. The structured array workload remained faster with encoding enabled, at 222.03 versus 241.25 us/eval.

An additional shape sweep found the encoded permutation path 14-16% slower at sizes 2 and 8, while singleton array, tuple, and record candidates still benefited. Candidate, result, and state digests matched across every comparison.

Validation

• uv run --extra test ruff check pyproject.toml src tests .github/scripts
• uv run --extra test ruff format --check pyproject.toml src tests .github/scripts
• uv run --extra test basedpyright src tests .github/scripts
• zuban check on all changed Python files
• uv run --extra test pytest tests -q (1607 passed)
• uv run --extra docs mkdocs build --strict