Group generated renderers into hash-assigned source files for incremental compilation

[gregjotau]

Summary

Renderers are generated into 32 source files chosen by a stable hash of the page-model name (<Registry>Part<n>), each with a holder class that owns the file's static resource. Renderers nest in their holder and reference only its STATIC; the registry is the only file that references every renderer. An unchanged file regenerates byte-identical output, so Gradle's incremental Java compilation recompiles one file plus the registry, and Kotlin's incremental compilation sees one changed Java source. KSP dependencies stay aggregating.

Measured on ReAI (344 renderers, warm daemon, --no-build-cache):

Task, one-template edit	0.11.2	This PR
:web-app:kspKotlin	3.0–4.0 s	3.6 s
:web-app:compileKotlin	1.0–1.2 s	0.09 s
:web-app:compileJava	6.1–6.3 s	0.3 s
Total build	10.4–12.1 s	4.3–4.4 s

Full :web-app:compileJava: 6.3 s → 2.8 s. Static resources: 1.6 MB → 2.8 MB (deduplicated per file instead of per module).

Correctness

• GoldenRenderTest: twelve renders (four example pages, three locales, form errors, 64-byte output buffer) recorded with the 0.11.2 compiler and compared byte for byte.
• All 105 tests pass. Consumer preflight: ReAI web-app and bedri (the latter compiles generated Java with -Werror).
• Upgrade transition from 0.11.2 with the build cache disabled: full build → this layout (incremental) → edit → retry → second edit → revert, all green.

What was tried and rejected

• One file per template: compileJava 0.27 s, but KSP grew 2.9 s → 4.5 s because it registers every generated Java file (Thim's own generation only grew by 0.23 s; Dependencies.ALL_FILES changed nothing), and resources grew to 6.5 MB.
• Package-private top-level renderer classes: javac's auxiliary-class lint rejects references from another file, fatal under -Werror.
• Keeping the ThimRenderer suffix: Gradle's previous-compilation data kept listing every renderer under ThimTemplates.java, so the first post-upgrade edit deleted all class files and retrying did not recover.

Details in PERFORMANCE_AUDIT.md.

🤖 Generated with Claude Code

[gregjotau]

Full edit loop on ReAI (:web-app:classes after touching one template, warm daemon, --no-build-cache, three repetitions each):

Task	0.11.2 single source	This branch
:web-app:kspKotlin	3.0–4.0 s	4.5–4.9 s
:web-app:compileKotlin	1.0–1.2 s	0.06–0.09 s
:web-app:compileJava	6.1–6.3 s	0.26–0.28 s
Total	10.4–12.1 s	5.2–5.6 s

Kotlin's incremental compilation also benefits, since only one Java file on its source path changes. KSP is now the dominant cost of a template edit, so the next step for the edit loop is inside the processor (property-lookup caching, or skipping regeneration for unchanged templates), not javac.

Upgrade transition verified with --no-build-cache: 0.11.2 full build → this branch (incremental, succeeds) → edit (succeeds, 2 real recompilations) → retry → second edit → revert, all green.