perf: Cache redundant work in transform phase

[danceratopz]

Requires #37

Summary

Add module-level and instance-level caching to seven hot paths identified via profiling. Each optimization eliminates redundant work that was repeated per-document or per-node during the build and transform phases.

What changed

1. Cache entry_points() in HTMLVisitor — Module-level _get_html_entry_points() replaces per-instance entry_points(group="docc.plugins.html") calls.
2. Cache entry_points() in Loader — Module-level _get_plugin_entry_points() replaces per-instance entry_points(group="docc.plugins") calls.
3. Cache dataclasses.fields() in PythonNode — Class-level _fields_cache ClassVar replaces per-call fields(self) in children and replace_child.
4. Cache _BoundsVisitor results in VerbatimVisitor — Instance-level _bounds_cache eliminates double tree traversal in enter()/exit().
5. Cache file lines in TextSource.line() — Lazy _lines_cache stores file content on first access, eliminating re-reads per line() call.
6. Cache Jinja2 environments — Module-level _get_jinja_env() (HTML) and _get_listing_env() (listing) replace per-render Environment(...) construction.
7. Share loaded renderers across HTMLVisitor instances — Module-level _LOADED_RENDERERS dict replaces per-instance self.renderers = {}, so EntryPoint.load() runs at most once per node type.

Why

The transform phase dominates runtime (~87% of wall time). Profiling showed that entry_points(), dataclasses.fields(), Jinja2 Environment construction, _BoundsVisitor traversal, file I/O in TextSource.line(), and EntryPoint.load() were all called thousands of times with identical arguments across documents. Caching these at the appropriate scope (module, class, or instance) eliminates the redundant work.

Test coverage

Each optimization includes three categories of tests:

• Behavioral tests — Verify the cached code path produces the same results as before.
• Call-count / spy tests — Use unittest.mock.patch to verify the expensive call happens only once when multiple consumers access the cache.
• Cache-keying tests — Verify the cache data structure contains entries with expected keys and value types.

Benchmark

Command:

hyperfine --warmup=1 --runs=5 --show-output 'taskset -c 0 uv run docc'

Individual optimizations

#	Optimization	Mean	Range	Speedup
—	Baseline	70.1s	67.9–74.4s	—
1	HTML entrypoints cache	24.0s	17.0–27.3s	2.9x
2	Loader entrypoints cache	48.0s	39.3–51.0s	1.5x
3	Dataclasses fields cache	49.4s	45.1–50.7s	1.4x
4	BoundsVisitor cache	50.2s	49.4–50.8s	1.4x
5	TextSource line cache	48.9s	48.3–49.4s	1.4x
6	Jinja2 environment cache	35.9s	29.6–38.2s	2.0x
7	Shared renderer cache	50.3s	49.5–51.4s	1.4x

Combined result

Variant	Mean	σ	Range	Speedup
Baseline	70.1s	2.6s	67.9–74.4s	—
All 7 cache opts	15.6s	0.4s	15.2–16.2s	4.5x

Per-phase breakdown (combined)

Phase	Baseline (median)	Optimized (median)	Change
Built	6.5s	6.6s	—
Transformed	61s	7s	-88%
Wrote	0.6s	0.6s	—

[danceratopz]

Ah nice, I didn't see that fc0d5eb had already landed, will rebase on main tomorrow 🥱