0.5.0

Performance and assurance release: the engine drops from 14.1 to
12.5 µs/formula (−11 %) on the full golden corpus — 3.3× the Rust
reference engine (41.4 µs same-day, same machine), with matrix/array-heavy
input a further −18 %. Layout output remains bit-exact across all
2 541 cross-engine fixtures. Test coverage now meets the release bar
across both targets (zero unmarked uncovered lines), and the entire
optimization diff passed an independent adversarial review with zero
findings. Rendering verified end-to-end on iOS (full simulator suite) and
macOS across SwiftUI, UIKit, and AppKit hosts.

Performance

• Parser hot state (MacroExpander stack/lexer/macro namespace,
  Parser.mode) opted out of dynamic exclusivity enforcement with
  @exclusivity(unchecked) — 14.5 → 13.2 µs/formula (−9 %).
  Single-threaded, non-aliased access verified per property.
• First-byte filters gate the macro-table and function-registry lookups,
  and a 128-entry direct-index table replaces dictionary hashing for
  single-ASCII-character symbols — 13.2 → 13.0 µs/formula, with a
  differential suite pinning fast-path == dictionary semantics.
• Two-layer macro namespace (shared read-only builtins + tiny per-parse
  layer, KaTeX's own design) stops a full CoW copy of the ~500-entry
  builtin table on every array/matrix environment — −18 % on
  environment-heavy input.
• Argument parsing no longer interpolates an "argument to '…'" error
  label on the happy path — 13.0 → 12.5 µs/formula (−3 %).
• The end-of-expression check is a first-byte-failing switch instead of
  a per-token Set<String> hash, and callFunction reuses the registry
  entry parseFunction already fetched (−1.7 %, interleaved A/B).

Testing

• Coverage raised to the release bar across both targets: zero unmarked
  uncovered executable lines in SwaTex and SwaTexRender (77 new
  tests, 817 total). Every intentionally uncovered line carries a
  grep-auditable INTENTIONALLY UNCOVERED proof marker. En route, the
  mhchem watchdog guard — long assumed dead code — was proven reachable
  and is now pinned by a test.
• The full performance diff was adversarially reviewed by an independent
  pass (aliasing/exclusivity, filter false-negatives, namespace scoping
  semantics, error-message parity): zero confirmed findings.