Support native trigonometric stages and checked canonical collection on Metal

[PraneethMerugu]

Summary

Complete the demonstrated LocalMath execution prerequisites for Potts authored
cell polarity and structured lifecycle transactions, without adding a second
executor or changing scientific publication laws.

• Admit native real sine calls through the existing closed unary-math analysis.
  Named and captured functions share the same selected-method ownership check;
  public method reflection identifies the native owner. Unsafe user overloads
  still undergo ordinary recursive analysis and cannot inherit native trust.
• Keep ascending/descending operand selection outside the canonical collection
  comparator, leaving one checked key-load call site in the existing bitonic
  loop. Comparator semantics, padding guards, barriers, pair writers, merge
  order, and duplicate validation remain unchanged.
• Add shared ordinary CPU/Metal contracts, nearest API and contributor docs,
  and a reproducible warmed public collection benchmark.

The collection regression exercises capacities 1, 2, 255, 256 and 257; reversed
mixed tuple keys; equal-key identity ordering; partial participation; closed
publication gates; and failure-atomic duplicate rejection. Trigonometric tests
compare actual stage publication with an independent higher-precision host
oracle and verify input preservation. They do not claim bitwise transcendental
equality, arbitrary array/complex support, or downstream lifecycle-product GPU
support.

Validation

• Complete ordinary package suite: 1,730/1,730, including package quality.
• Full strict documentation build: passed, with doctests and exported API
  coverage. Documenter reports a non-fatal missing navbar repository-link warning.
• Focused collection CPU and actual Metal contracts: 65/65 each, bounds checked.
• Complete joined actual Metal inventory: 416/416, including cross-domain
  scientific witnesses, with --check-bounds=yes and scalar indexing disabled.
• Joined CPU cross-domain scientific inventory: 29/29, including compacted and
  authored-domain examples, with bounds checking enabled.
• Final combined independent review: no blocking findings across all 15 changed
  files, including comparator ordering, native-method provenance, test inventories,
  and bounded documentation claims. git diff --check passed.

Local warm Metal medians at 256/4096 records were 0.846/1.313 ms before and
0.765/1.241 ms after, using three warmups and nine batches of ten submissions.
Ranges overlap: no general speedup or machine-independent threshold is claimed.

This is the demonstrated G04 LocalMath execution-prerequisite companion in the
ecosystem delivery map. The separate Core nested lifecycle numeric-conversion
failure remains under investigation; this PR does not claim to solve it.