0.13.0

• Registry dispatch stays JIT-compilable. The generated TemplateSet.render kept the index switch and the subclass fallback chain in one method. Above about 215 templates that method exceeded HotSpot's 8000-byte huge-method limit and was never compiled. ReAI web-app (353 templates) paid about 190 ns per render instead of about 10 ns. Dispatch now uses a small dispatcher, 256-template chunk methods and a separate fallback. Rendered output is unchanged. (#64)
• Removed HtmlOutput.url(TrustedUrl). Nothing called it, and calling it skipped context-path resolution. (#64)
• The compiler builds its naming regexes once, and stale docs were fixed. (#64)
• KSP 2.3.12, Spring Framework 7.0.9, Spring Boot 4.1.1, JUnit 6.1.3. (#65)

0.12.2

Small compiler improvements with unchanged generated output and runtime behavior:

• Generate Java at its final indentation, eliminating whole-renderer copying and regex formatting.
• Reuse route matches within one compilation while checking HTTP methods, enum coverage and diagnostics for every use.

On ReAI/M5 Max, median complete rebuilds after different-length static template edits improved from 6.68 s to 6.15 s; same-length resource-only edits improved from 5.60 s to 5.53 s. Results vary by edit and machine. No separate development compiler or additional generated resources.

Validated with the full clean build, isolated-project checks, golden renders, route-cache regressions and byte-for-byte consumer generated-output comparisons. See PR #63 and PERFORMANCE_AUDIT.md for measurements and limits.

Thim 0.12.1

What's Changed

• Cache repeated property discovery during template compilation by @gregjotau in #61
• Release 0.12.1 by @gregjotau in #62

Full Changelog: v0.12.0...v0.12.1

Thim 0.12.0

What's Changed

• Group generated renderers into hash-assigned source files for incremental compilation by @gregjotau in #59
• Release 0.12.0 by @gregjotau in #60

Full Changelog: v0.11.2...v0.12.0

Thim 0.11.2

• Dispatch generated template registries through a class index instead of linear == and
  instanceof chains. With 422 page models the last model cost about 1.1 µs per request;
  every model now resolves in a few nanoseconds. Subclasses of open models still fall back
  to the previous instanceof chain. The new registry also compiles faster: ReAI's
  compileJava for 344 renderers fell from 10.8 s to 6.3 s.
• Write the message usage manifest whenever a catalog exists, so generateMessages=false
  keeps build-wide dead-key detection. A module can now render templates from a catalog
  whose typed factories another module generates instead of compiling a second factory class.
  The module-local unused check that previously ran at compile time is replaced by the
  build-wide thimMessageUsageCheck.
• Record the measurements behind these changes, and the buffer and text-encoding ideas
  that were rejected, in the performance audit.

Thim 0.11.0

Thim 0.11 adds the no.beint.thim.settings plugin and supports Gradle Isolated Projects for Kotlin compilation and build-wide template, message and CSS validation.

Add the settings plugin alongside the existing project plugin, both at version 0.11.0. KSP is upgraded to 2.3.11; set ksp.project.isolation.enabled=true so its generated-source wiring works with Gradle's --isolated-projects flag. See the README for complete migration commands.

• Settings and project plugins use separate artifacts to keep KSP off the settings classpath.
• Validation collects declared project artifacts, including non-Thim modules, without cross-project configuration.
• Compiler cache inputs use relocatable paths. Template strictness and unused-message/CSS validation are preserved.
• The full library build passes 102 tests, including Kotlin consumer compilation and configuration-cache reuse under isolation.

All five libraries/plugins and both plugin markers are published to Maven Central. Spring Boot's separate isolated-packaging limitation remains upstream; this release does not justify enabling isolation globally for Boot applications or claim a 30% build-time improvement from isolation alone.

Thim 0.10.2

Constant translations now return their compiled string literal, avoiding a new string on every lookup. The runtime and Spring adapter implementations are unchanged.

• Direct constant-message benchmarks drop from 48 allocated bytes per lookup to effectively zero.
• ReAI's generated message Java is 12.1% smaller and its compiled message classes are 5.9% smaller.
• Locale and regional fallback, references, escaping, empty/Unicode text, parameters, plural/select handling, and null-locale validation retain their behavior.

Validation: 100 library tests and clean builds of ReAI, Utin, Eteo, and Ecomtools. Packaged comparisons match for 399,882 message resolutions, 122,724 reference resolutions, 57,126 null-locale checks, and 24 HTML renderings per version. Runtime and Spring class files are byte-identical to 0.10.1.

Benchmark scope and raw results are documented in PERFORMANCE_AUDIT.md. These are operation and generated-size improvements; no whole-application latency or build-time speedup is claimed.

Thim 0.10.1

Thim 0.10.1 reduces template compilation work and speeds up integer output while preserving the runtime and Spring APIs.

• Indexed source locations: parsing ReAI's 388-template corpus measured about 21× faster.
• Fragment expansion: 30% less time and 64% less allocation on the measured subset.
• Direct integer formatting: about 2× faster, without temporary allocations.
• Reuse opening-tag source locations, validate HTMX 4 QUERY routes, and align Kotlin bytecode with the documented JDK 26 requirement.

Compatibility validation: 98 library tests pass; clean builds pass in ReAI, Utin, Eteo, and Ecomtools. All 41 generated files are byte-identical to their 0.10.0 baselines, and 24 packaged renderer cases produce identical HTML with old and new runtime/Spring jars. HTTP smoke checks cover localization, UTF-8, content lengths, form errors, and escaping.

The measured gains describe specific operations, not total application build or request time. See PERFORMANCE_AUDIT.md for details.