feat(skills): add meeting-prep

[jdjioe5-cpu]

Summary

• add the `meeting-prep` graph skill
• emit a bounded `runx.meeting.brief.v1` packet (agenda, decisions, risks, questions, follow_ups, citations) without inventing attendee history
• mark missing or private context explicitly in `missing_context` instead of fabricating it
• add one sealed bounded-brief fixture and one `needs_agent` stop fixture for insufficient context
• include the local harness result used for submission evidence

What this skill does

The skill reads four pieces of evidence:

• `event{id, title, start_at, duration_minutes, attendees[]}`
• `attendee_notes{attendee_id: notes}` keyed by attendee id from the event
• `thread_snippets[{thread_id, author, sent_at, body}]` prior conversation rows
• `public_links[{url, fetched_at, digest, excerpt}]` fetched content the operator explicitly chose to share

It verifies that every cited snippet, attendee note, and public link digest actually appears in the inputs, and that no `attendee_history`, `mail`, or `calendar` field has been supplied as private. The packet step is an agent-mediated judgment, so an unattended run stops at `needs_agent` instead of fabricating the brief.

Verification

• `runx-cli 0.6.14`
• `runx harness ./skills/meeting-prep --json`
• 2 cases passed, 0 assertion errors
• receipt `sha256:82c6c425e68c7a1e23f4260a609a756447c5b5aed9af40c5c18a23a639fc0916`

Why a real operator would use it

Meeting prep is a high-frequency operator workflow and a good test of scoped context: the skill receives bounded inputs and returns a brief without claiming access to anything it was not given. It is the canonical pattern for bounded-context skills in the runx catalog.

Refs Frantic Bounty #27 (runx skill: meeting prep from bounded context, USD 10, x402 rail on Base).