feat(FN-3377): add llama.cpp provider integration with pi extension and mob

This merge adds a new llama.cpp local AI provider integration, including a `pi-llama-cpp` CLI extension package with model resolution and context retrieval, a `LlamaCppProviderCard` UI component for provider configuration, backend API routes for model registration and llama.cpp probe/health checking

Fusion-Task-Id: FN-3377

.changeset/fix-issue-33-dockerode-missing.md

@@ -0,0 +1,5 @@
+---
+"@runfusion/fusion": patch
+---
+
+Harden the publish path against dockerode-class missing-dependency regressions (#33). Adds a generalized invariant test that walks `tsup.config.ts` and asserts every non-builtin `external` is either a runtime dep or in an explicit transitive-allowlist, plus a pre-publish smoke step in `pnpm release` that packs the public tarballs, installs them with plain `npm` into a clean temp dir, and invokes the bin — catching the dockerode-class bug (and others like missing `files` globs) before publish, since pnpm hoisting masks it in the workspace.

.changeset/launcher-update-notice.md

@@ -0,0 +1,5 @@
+---
+"runfusion.ai": patch
+---
+
+Surface "new version available" notices to users who run `npx runfusion.ai` without ever opening the dashboard. The launcher now reads the existing `~/.fusion/update-check.json` cache (written by the dashboard's update-check service) and prints a one-line stderr notice when a newer Fusion is published. When the cache is missing or older than 24h, a fire-and-forget fetch against the npm registry refreshes it (1.5s timeout) so non-dashboard users still pick up updates on their next run. Disable with `FUSION_NO_UPDATE_CHECK=1`; auto-skipped in CI and non-TTY contexts.