ci: add Release Please and auto-publish the compiled .dict on release

[laurigates]

Implements the release automation the README's CI/CD section already promised ("On merge, the compiled .dict is published as a GitHub Release artifact at a stable URL") and that assemble.yml's comment referenced as the "release-dict workflow (not yet authored)".

What this adds

Release Please (release-please-config.json + .release-please-manifest.json)

• release-type: simple — maintains a release PR from Conventional Commits on main; merging it tags the commit, updates CHANGELOG.md, and creates a GitHub Release.
• Keeps the workspace version in Cargo.toml in sync via a generic updater annotation (version = "0.0.1" # x-release-please-version).

release.yml — canonical two-job pattern:

• release-please opens/merges the release PR and exposes release_created / tag_name outputs.
• publish-dict (gated on release_created == 'true') installs libvoikko + voikko-fi + JDK + just, runs just generate (download corpus + jar → assemble → compile), then attaches puhekieli_fi-<tag>.dict and niinku-<tag>.combined to the release via gh release upload.

Docs — README CI/CD section + repo layout updated; README "Quick start" now points users at the Releases page for pre-built .dict files; the stale assemble.yml comment is corrected.

Why simple and not rust

The crates aren't published to crates.io and use an inherited [workspace.package].version, which the rust plugin handles poorly. simple cleanly owns the tag/changelog/release, and the generic annotation keeps Cargo.toml accurate. The real deliverable is the .dict, not crate publishing.

Why publish-dict is a dependent job, not on: release

Releases created with the default GITHUB_TOKEN don't trigger downstream workflows (e.g. on: release: published). Running publish-dict as a needs:-dependent job in the same run avoids that pitfall — no PAT required.

Verification

I ran the workflow's exact build path locally (replicating just generate, since just isn't on this box):

• Downloaded OpenSubtitles fi_50k (50,000 lines) and dicttool_aosp.jar — network + URLs good.
• assemble: curated (169) + OpenSubtitles (50,000) → Voikko filter → 7,550 puhekieli tokens written to .combined; example forms (miksköhän, saaks, oliks, niinku, mä) all present.
• compile: dicttool produced a 46 KB puhekieli_fi.dict. ✅
• cargo fmt --check, cargo test --workspace (36 tests), config JSON, and workflow YAML all validate; cargo metadata confirms the inline Cargo.toml comment parses.

Downloaded corpus/jar and built artifacts are already covered by .gitignore (/data/cached/*, /data/out, *.dict, /dist, /tools/*.jar), so nothing build-generated is committed.

Notes / follow-ups

• The release PR that Release Please opens won't itself trigger ci.yml (same GITHUB_TOKEN limitation). If you want CI to run on the release PR, swap the action's token for a PAT/GitHub App. Left as-is for now since it needs a secret.
• First release from 0.0.1: a feat: will propose 0.1.0 (pre-1.0 default). Adjust later via config if you'd rather stay on 0.0.x.

https://claude.ai/code/session_018xYxT7WXd64cm6C2wSbkFq

---

Generated by Claude Code