feat(skills): port Hermes /learn — author skills from sources

[TheSmokeDev]

Add a user-initiated, source-driven skill authoring command, porting
Hermes Agent's /learn. Point it at a URL, a local dir/file, the current
conversation, or pasted notes; it gathers the material, distills a
standards-compliant SKILL.md, and stages an inert draft under
skills/generated/ that graduates through the existing /skills gate.

Model-agnostic by construction: the only LLM call is reasoning_step
(routed through the runtime lane system, whatever /model selected);
source gathering uses framework primitives (httpx+trafilatura, stdlib
file reads, the session transcript) — never a model-native tool surface.

Reuses every existing rail — write_skill, skill_guard.scan_skill, the
usage sidecar, and the operator-gated promotion flow — adding no new
storage, security, or promotion path. A /learn draft seeds the reuse
counter to the eligibility threshold so explicit operator intent stands
in for auto-capture recurrence evidence and the draft surfaces in
/skills review.

• cognition/skill_learn.py: parse/gather/distill/orchestrate core
• cognition/skills.py: optional SkillSpec.body rendered by write_skill
  (back-compatible; empty body keeps the auto-capture stub)
• core_handlers.py + commands.py: wire /learn (router, operator)
• tests/test_skill_learn.py: 14 tests, reasoning_step mocked

Co-Authored-By: Claude Opus 4.8 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01PuH3Z3zEyjkMCpywRuaqgj