v1.4.0 — chat→skill: turn a live session into a replayable skill

Major agent primitive. Mid-chat, an agent or user can now distill a tool-trace into a reusable skill. Two kinds:

• HARD — deterministic recorded.json replay. Draft → review → commit. Zero further LLM calls at replay time.
• SOFT — guidance.md retrieved on demand via list_skills/run_skill. Great for softer heuristics that don't want to be pinned to exact taps.

Exposes new agent tools draft_skill + save_skill, new REST endpoints /api/skills/draft-from-chat + /api/skills/save-from-chat, chat-loop intercept across every native + on-device provider, Claude Code MCP parity, compat matrix + verifier extended to both kinds, and frontend Skill Hub badges + guidance viewer + 'Save as skill' modal.

717 python tests pass on this branch (up from 674 pre-feature). Full 8-milestone breakdown in the PR description.

pip install --upgrade ghost-in-the-droid==1.4.0

Contributor: @C-K-Loan (core-harness work); community + iOS PRs from v1.3.0 line remain the shipped baseline.