feat(mcp): compact job waiting and evidence delivery

[tulayha]

Summary

• add compact job status and bounded wait_job observation contracts for HTTP and MCP
• keep full machine job records separate from model-visible evidence content
• expose concise completed-job evidence through get_job_evidence and bounded follow-up materialization
• return the existing local path, HTTPS URL, or MCP resource URI for boards, frames, and clips so agents can actually present them
• make the search_moments and query_video input fields explicit in the public skill and tool description
• treat per-artifact evidence rendering as one observable wait phase and default waits to 30 seconds
• remove polling cadence from compact long-poll responses
• trim the README agent section back to a product-facing overview

Why

The previous agent flow repeatedly fetched full job results while waiting, and structured MCP output could hide inline images from Codex and Claude. Evidence-board rendering also changed the raw progress stage for every artifact, waking the agent repeatedly during a single search. The projected delivery target was then discarded when concise evidence text was composed, allowing a run to finish with an unlinked “View evidence board” label despite the board existing.

Result

Agents can wait on compact tokenized status, fetch completed evidence once, inspect boards and selected follow-up artifacts, and present an actual local image path, HTTPS download, or MCP resource. Raw diagnostic progress remains available while internal render/validate/publish transitions no longer cause separate wakeups.

Validation

• MCP suite: 48 passed, including 7 subtests
• focused local/resource/HTTPS evidence projections: 3 passed, including 3 subtests
• real installed stdio MCP returned the existing Hermione board as text + inline image + resource link; its reported local JPEG path exists
• repository and global VidXP evidence skills validated
• Ruff and git diff --check passed