refactor: reuse pattern execution contexts

[chubes4]

Summary

• construct the normal, runtime-DOM-restricted, and probe PatternContext graphs once per HtmlTransformer
• reuse those immutable service graphs across recognition attempts and sequential transforms
• keep per-transform data in HtmlTransformerSession; cached closures continue resolving the current session through the transformer

This is a bounded maintainability/performance slice from #242. It removes repeated allocation of the full callback graph and PatternRecursiveConverter on every recognizePatterns() call while preserving the explicit normal/restricted/probe policy boundaries.

Behavior and performance proof

• fresh and reused transformer instances produce identical patterned output across sequential transforms
• staged dispatch and probe contracts remain green
• five alternating baseline/candidate benchmark pairs over 1,500 nested-container transformations were faster in every pair
• median elapsed time improved from 1319.6 ms to 1256.8 ms (4.8%)

The timing benchmark is external evidence rather than a brittle CI threshold.

Verification

• composer test
• 280 parity fixtures
• package install proof

AI assistance

OpenAI GPT-5.6 Sol via OpenCode was used to trace context construction frequency and session capture semantics, implement constructor-owned context reuse, add reused-instance behavior coverage, benchmark against the merged baseline, and run verification. Chris Huber is responsible for every line.