fix(ci): isolate rust-cache per target to prevent GLIBC mismatch

Add matrix.target as cache key for Swatinem/rust-cache so aarch64-musl
cross-build doesn't reuse host-compiled build scripts with newer glibc.
Also update AGENTS.md git workflow to reflect main-only branch strategy.

chore: Release clawhive-schema version 0.1.0-alpha.20

chore: Release clawhive-schema version 0.1.0-alpha.19

chore: update Cargo.lock

Merge branch 'dev'

chore: Release clawhive-schema version 0.1.0-alpha.16

feat: chat-based skill install system

Add /skill analyze, /skill install, /skill confirm slash commands for
installing skills directly from chat (Discord, Telegram, TUI).
Includes security analysis with risk scoring, two-step confirmation flow
with time-limited tokens, and interactive confirm/cancel buttons.

Other changes:
- Auto-normalize GitHub URLs (tree/blob/repo → archive/raw download)
- Inactivity gap markers injected into session history for LLM context
- Show actual error messages to users instead of generic "Internal error"
- /skill usage hints when subcommand arguments are missing
- HTML download guard for skill install (reject non-archive responses)
- Clear expired session JSONL on session reset

feat: detect bot role mentions in Discord channels

Add role-mention detection so the bot responds when mentioned via its
managed role (@botName App), not only via direct user mentions.
Also add network permission redesign plan doc.

feat: clickable approval buttons for Discord and Telegram

Replace text-based /approve instructions with interactive UI:
- Discord: Message Components (3 buttons: Allow Once, Always Allow, Deny)
- Telegram: Inline Keyboard with same 3 options
- Button clicks generate synthetic /approve commands, reusing existing logic
- Text /approve command preserved as fallback
- New DeliverApprovalRequest bus message for structured approval delivery

v0.1.0-alpha.12

🐛 Fixes
- Guard against empty Discord replies when upstream API returns errors
- Include full diagnostic details (message, code, extra fields) in ChatGPT/Azure OpenAI SSE error logs instead of "unknown error"