v0.2.1

• Escape captured dynamic end/while delimiters like TextMate, including spaces,
  #, and hyphens. This fixes extended-mode delimiters being interpreted as
  regex syntax and captured hyphens becoming character-class ranges.
• Preserve vertical tabs and non-ASCII whitespace as literals in extended-mode
  regexes, matching Oniguruma's separator rules.
• Report degraded highlighting when a captured dynamic delimiter exceeds the
  existing substitution limit, including cached and incremental results.
• Release the source style table after C token-buffer conversion, reducing
  memory retained by C and C++ results.
• Add oracle-backed Unicode, mixed-line-ending, and truncated-input recovery
  checks for every B/C promotion-group language, covering bundled, cached, and
  uncached incremental tokenization.
• Match TextMate's literal injection-selector semantics for asterisks, fixing
  missing invalid-angle-bracket scopes in Angular HTML attributes.
• Include dependencies reached through rule-local repositories in generated
  bundles, restoring embedded Python highlighting in Wikitext syntaxhighlight
  blocks and retaining those dependencies in subset bundles. Loading the
  previously omitted grammars increases bundled Wikitext construction plus
  tokenization from 2.62 to 4.82 ms on the catalog fixture (median of five
  alternating runs, 30 iterations each).
• Reject incomplete tokenization in the cold-throughput benchmark.
• Reduce cached-line overhead by avoiding unused source-text/state clones and
  regex scratch initialization on cache hits. Across the 160 B/C languages,
  five alternating process-cold sweeps of 30 iterations using source-asset
  closures measured 36.27 → 36.85 MB/s aggregate (+1.6%), with unchanged
  benchmark scope digests.
• Reduce uncached regex work by reusing captures already computed during
  candidate selection, skipping duplicate candidate guards, and compiling
  optional groups as ordered choices instead of counted loops. Preserve
  capture rollback, lazy and possessive matching, and Unicode scope ranges.
  Seven alternating steady-state samples on a Ryzen 9 9950X measured 1.5–2.9%
  higher HTML throughput on the Markdown, TypeScript, C++, Rust, and C stress
  fixtures, with line caching disabled and unchanged scope/HTML digests.