use knowledge base skill for incoming voice transcripts

[max-tet]

Instead of using the shipped prompt for incoming pebble voice transcripts, use the voice-command skill in the knowledge_base. The System prompt should just assemble the parts but not provide much context itself.

Aim: evolving the behavior should be quick and easy through the knowledge_base.

[ClaydeCode]

Done — PR #79.

What I did:

• Created src/clayde/skills_builtin/voice-command.md with all the detailed voice-command handling instructions (speech-to-text error handling, KB working path, phonetic disambiguation example). This is a proper skill file shipped with the image.
• Slimmed _SYSTEM_PROMPT_TEMPLATE down to minimal structural glue: identity line, the skills catalog, skill-usage policy, and the JSON output contract. No behavioral context is hardcoded in Python anymore.
• Fixed discover_skills() so non-builtin skills (personal/shared) take priority over builtin on name collision — this lets you override the shipped voice-command skill by dropping your own version in ~/skills/personal/voice-command.md.

Evolving the voice-command behavior now means editing a markdown file (or putting your own override in your personal skills directory) — no code rebuild needed.

[ClaydeCode]

Created src/clayde/skills_builtin/voice-command.md containing all detailed voice-command handling instructions (speech-to-text handling, KB path, phonetic disambiguation). Slimmed down the system prompt template to minimal structural glue (identity, skills catalog, JSON contract). Fixed discover_skills() to give non-builtin skills priority over builtins on name collision, so users can override the shipped voice-command skill with a personal version. Updated tests to match. All 293 tests pass. Opened PR #79 closing issue #78.