perf: use typed attributes in extraction hot paths

[ryanfowler]

Summary

• Use cached typed attribute keys for fixed names in scoring, cleanup, normalization, metadata extraction, and semantic compilation.
• Add lookup-only keys for aria-live and legacy data-math attributes.
• Preserve local-name, namespace, case, and duplicate-order behavior.

Performance

perf showed repeated dynamic attribute-name classification in whole-document scans. The focused changes remove that work from the raw HTML to Markdown path. Criterion comparisons against the saved baseline show these median changes:

Benchmark	Before	After	Change
smoke medium extract + Markdown	8.477 ms	7.973 ms	-11.46%
pipeline medium reference	10.535 ms	10.250 ms	-2.71%
pipeline medium code	9.425 ms	8.960 ms	-4.93%
pipeline medium math	8.645 ms	8.421 ms	-2.59%
pipeline medium tables	9.172 ms	8.850 ms	-3.51%
pipeline large ordinary inline	77.068 ms	75.189 ms	-2.44%

Lazy Markdown rendering had no significant change. This result is expected because the change targets parsing, extraction, and lowering.

Validation

• cargo fmt --check
• cargo fmt --manifest-path cli/Cargo.toml --check
• cargo clippy --all-targets --all-features -- -D warnings
• cargo clippy --manifest-path cli/Cargo.toml --all-targets -- -D warnings
• cargo test
  • 621 unit tests
  • 218 snapshot fixtures
  • 14 public API tests
  • 130 Mozilla Readability compatibility tests
  • 3 doctests
• Two independent review rounds. The first found one missed scoring lookup. The second found no actionable issues.