libmorpheus 0.3.0

Version 0.3.0 is a compatible feature release. It adds Greek lemma generation, a generation-options structure, an owned generation-result type, generation accessors, and a result-limit status. It does not remove a public function, change a published signature, reassign an existing constant, or change the layout of morpheus_analysis. ABI 2 and SONAME 1 therefore remain correct.

The new C generation surface and Deno generate()/generateRaw() methods are experimental. That maturity label does not weaken the ownership, symbol, or binary-compatibility contracts of the additive ABI 2 declarations.

Highlights

• Build a deterministic, validated, duplicate-preserving reverse generation index from the pinned Greek stem sources.
• Expand derivations and continuations offline without reviving the historical gener and do_conj request paths.
• Generate through a context-local internal service that invokes only GenStemForms() and GenIrregForm().
• Normalize generated interpretations at the MPL/AGPL bridge, preserving dialects, duals, duplicates, stable ordering, and explicit result limits.
• Publish the additive C API and nonblocking typed Deno binding.
• Qualify cold and warm generation, small and maximal paradigms, concurrency, and memory behavior through benchmark schema 2.
• Improve the standalone Deno and repository documentation, including Docker, stemlib provenance, FFI permissions, and the experimental maturity notice.

Compatibility

• Project version: 0.3.0
• C ABI: 2
• SONAME major: 1
• Previous release tag: v0.2.0 at 5c071548a7ceea05340dffc60591567cff7ca6c5

Licensing boundary

The inherited generator, source-format translation, generation service, normalizer, derivation expansion, and historical compatibility expressions remain MPL-2.0. The independently written public API orchestration, Deno binding, tests, build and release support, and documentation remain AGPL-3.0-or-later. Generated index data follows the licensing and redistribution terms of its stemlib inputs.

Distribution contract

The candidate continues to produce data-free native archives and SHA-256 files for Linux x86-64 glibc, Linux aarch64 glibc, and macOS arm64, plus the platform-independent libmorpheus-deno-0.3.0.tar.gz source archive. Native binaries and stem data are not embedded in the Deno archive.

The Deno source package is configured for JSR under the reserved name @humanities/libmorpheus at candidate version 0.3.0. Its metadata, exact source-only contents, import command, and strict registry dry-run are covered by the ordinary Linux CI. It includes the public /data and /native commands and their internal support, but no downloaded binary or linguistic data. The JSR package retains the AGPL notice and the documented MPL boundary to the separately distributed native runtime. The /native command verifies and extracts the matching release asset without a C toolchain.

Container builds remain qualification artifacts. They embed the pinned Alpheios stemlib and its validated derived gener.index so Docker qualification covers real experimental generation. Such an image must not be published until the dataset and derived-index redistribution terms have been confirmed.

Benchmark evidence

The accepted schema 2 report was produced on Apple Silicon from 66e656d27aedce04cc7b96fe84aff3aca8a35afd with Deno 2.9.5, Apple Clang 21, the pinned Alpheios revision, and the canonical generation-index digest. Its
SHA-256 is 82875bed7e33312a1315819dc8c73fb53aa79a05c8aef0407b5768c2baf2a290. The release validator accepts every required measurement and identity.

The commits between that measured revision and this finalization change only JSR data/native acquisition, the opt-in Emscripten preparer build, Docker and release orchestration, licensing, and documentation. They do not change src/, include/, bindings/deno/mod.ts, bench/compare.ts, or test/fixture.json; the measured analysis and generation paths are therefore unchanged. The report is retained as preparation evidence rather than described as tag-exact evidence.

Compared with 0.2.0, analysis throughput changes by approximately -1.2%, -1.5% and +0.6% for one, two and four FFI contexts. Persistent cruncher throughput changes by +0.6% and cold startup throughput by -2.2%. Peak FFI RSS increases by approximately 6.6% to 7.5%, while measured RSS growth decreases. These small changes are accepted. Schema 2 establishes the initial generation baseline: warm lo/gos generation scales from about 8,006 to 22,965 operations/second, while the maximal i(/hmi workload scales from about 132 to 175 operations/second across one to four contexts.