FIX: Make RenderContext parent-queries answer is_a? semantics

[gschlager]

Summary

Two RenderContext changes, second motivated by benchmarking the first:

1. Make parent-queries answer is_a? semantics. RenderContext#has_parent?, #find_parent, and #count_parents were keyed on exact class via parent_cache, so a user-defined subclass like class CustomUrl < AST::Url slipped past has_parent?(AST::Url). That's surprising — callers reasonably expect "do I have a Url ancestor" to include subclasses — and it makes built-in checks like the renderer's in_link_label? (which drives label-aware ] escaping) silently miss custom AST nodes. The three predicates now use is_a? semantics.

2. Drop the parent_cache. After (1), the cache was only earning its keep on has_parent?'s exact-class fast-path; find_parent and count_parents already had to scan @parents to answer polymorphic queries. Benchmarking showed the cache cost more than it saved — the linear scan over realistic parent depths (3–8) is faster than maintaining the hash. The class loses the parent_cache: kwarg, build_cache, and the incremental cache-update logic in with_parent.

Benchmarks

End-to-end (bench/bench.rb --isolated, YJIT, Ruby 3.4.8):

Report	Baseline (with cache)	No cache	Delta
simple	36.8k/s	39.3k/s	+6.7%
nested	38.3k/s	41.4k/s	+8.2%
list	35.3k/s	38.4k/s	+8.6%
table	11.3k/s	12.1k/s	+6.9%
quote_nested	26.0k/s	27.7k/s	+6.4%
code	83.7k/s	88.9k/s	+6.1%
url	44.9k/s	47.0k/s	+4.8%
mixed	13.1k/s	14.0k/s	+6.7%
large_doc	744/s	797/s	+7.2%
escape_plain	269.1k/s	268.6k/s	-0.2% (noise; escaper-only path)
escape_mixed	5.65k/s	5.67k/s	+0.3% (noise; escaper-only path)

Micro-bench (RenderContext push + query at depth 3 / 8):

• with_parent push: 2.1× faster — no hash dup, no array allocation
• has_parent? exact hit: -15% (43 ns vs 36 ns)
• has_parent? miss: +9–12% faster — the key? probe was strictly extra work before the scan ran anyway
• find_parent / count_parents: unchanged (already scanned)

Real workloads hit the miss path more than the hit path (has_parent?(Code) / (Url) / (Email) on every Text node — most text isn't inside those), so the miss-side win dominates. with_parent's allocation savings compound across every nested element.

Test plan

• bundle exec rspec — 3122 examples, 0 failures
• bin/mutant run on all RenderContext methods plus the now-dependent renderer/escape subjects — 100% mutation coverage (93/93 + 391/391 kills)
• bin/lint — no offenses
• Benchmarks above

Related

• FIX: Escape ] in URL label to prevent early link termination #40 (the original ]-escape fix)
• FIX: Move ] label-escape into MarkdownEscaper so image-in-link works #41 (moved the escape into MarkdownEscaper; needs this PR to correctly trigger for Url subclasses)