Add skill system for shareable prompt-level agent behaviors

Skills are TOML manifests containing instructions injected into the LLM
context. They can be loaded from GitHub repos, URLs, or local files and
activated via /skill commands or custom slash commands from any channel.

Five defense layers protect against malicious skills: static analysis
(Aho-Corasick + skill-specific regex), hard tool whitelist (registry +
execution level), resource restrictions (domains, workspace paths, tool
call budget), BLAKE3 approval pinning, and structural prompt isolation
via <external_skill> wrapper with reassertion block.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

Cargo.lock

@@ -2157,6 +2157,7 @@ dependencies = [
  "tokio-postgres",
  "tokio-stream",
  "tokio-test",
+ "toml",
  "tower",
  "tower-http",
  "tracing",