Release v2.5.1 — terser preservation fix for window.__hyperfixi_i18n · codetalcott/hyperfixi

[2.5.1] - 2026-05-24

A single-bug patch for a v2.5.0 publishing regression that broke every localized htmx attribute.

hyperfixi-hx.js / hyperfixi-hx-v4.js: window.__hyperfixi_i18n lost to terser (@hyperfixi/core). The Phase 8 orchestrator's public-API singleton (window.__hyperfixi_i18n = { register }) was being eliminated by terser in the published v2.5.0 hybrid-hx and hybrid-hx-v4 bundles, silently breaking every vocab/htmx/{lang}.js module on load with "loaded before the htmx-compat orchestrator". Two layers fixed: (1) the module-level installPublicAPI() invocation is now exported and called explicitly from each bundle entry that includes htmx-compat, so terser's unused: true, toplevel: true pass can't drop it; (2) the API is exposed via bracket-access (window['__hyperfixi_i18n']) to bypass terser's properties.regex: /^_/ mangling. Localized htmx attribute names (hx-obtener, hx-取得, hx-احصل, …) now resolve correctly in the published bundles, matching their already-working behavior under the source-aliased vitest suite.

Orchestrator public-API gate in pre-publish-check + release-smoke. New 2-test Playwright spec (packages/core/src/compatibility/browser-tests/i18n-orchestrator-api.spec.ts, ~600 ms) asserts typeof window.__hyperfixi_i18n.register === 'function' in both shipped htmx bundles. Wired into .github/workflows/pre-publish-check.yml (fails the workflow on regression) and the --matrix stage of examples/release-smoke/run.mjs (so the gate fires against the registry-installed tarball, not only the locally-built dist). Surgical — no swap-pipeline dependency. The broader i18n-htmx.spec.ts remains un-wired pending two pre-existing bugs unrelated to this release: a reactivity / notify-hook bug breaking localized hx-live re-renders, and a swap-pipeline bug stringifying DocumentFragments in fetch ... as html → put it into target.