fix(mcp): complete follow-up evidence delivery

[tulayha]

Summary

• repair create_evidence_clip for the control-plane application used by stdio and Streamable HTTP
• add materialize_job_evidence for explicit follow-up keyframe/clip batches of up to 10 evidence IDs while retaining the automatic 5-item search/query ceiling
• reuse the existing evidence renderer and transport-specific artifact projection rather than requiring timestamps or rerunning retrieval
• update the public VidXP skills to surface returned evidence and accurately describe the current anonymous actor-cluster scope
• default the ingestion skill to dialogue and scene, leaving actor indexing opt-in until its exact detection and overlay operations are available through MCP

Validation

• 74 focused tests passed, plus 7 subtests
• Ruff and git diff --check passed
• both public skills passed quick_validate.py
• real stdio MCP regression verified the formerly crashing evidence-clip path
• real MCP materialization returned a complete 10-item evidence batch
• shared server schema verified 10 follow-up items and the retained 5-item automatic ceiling

BEGIN_COMMIT_OVERRIDE
fix(mcp): restore clip creation and 10-item evidence delivery
END_COMMIT_OVERRIDE