ci(target): gate MCP evals in Modal

[walkerhughes]

Summary

• add four deterministic Harbor tasks for Target search, TCIN detail, pickup availability, and shopping-list pricing
• score answer correctness independently from expected MCP use and endpoint bypasses
• run the full Claude Code agent loop in Modal
• pin Harbor 0.18.0 and install the exact PR SHA in each eval image

Validation

• make check
• make coverage (89.22%)
• 28 process scenarios and 16 outcome scenarios across four evals
• Harbor Modal configuration parse
• official plugin and skill validators
• workflow YAML parse

Stack

Depends on #64, which depends on #63.