Add MCP server agent evals; require SCAPI/CIP skill reads

[clavery]

Summary

Adds agent-level evals for the MCP server and the tool/guidance fixes they surfaced.

Evals (packages/b2c-dx-mcp/evals) — a claude plugin eval suite that runs the locally built server:

• Compliance (default, pnpm run mcp:eval): every case, plugin loaded, --ablation none. Checks tool choice, required skill reads, tools that must not be called, and no terminal fallback.
• Enhancement (pnpm run mcp:eval:ablation): ablation-tagged docs cases run with and without the plugin and report Δ. Path graders are arm: with-only so the no-plugin arm is judged on its answer.
• Prompts are phrased as a developer new to B2C Commerce would ask, with no tool hints.
• Offline only today (tier0); fixtures use a fake dw.json. docs_schema_read is flagged (whole-XSD output is too large) and its case is in disabled/.

MCP fixes

• skillRead is now a required true literal on scapi_search, scapi_execute and cip_query. Before, only Opus 5.5 read the SCAPI skill first; Sonnet 5/5.5 now do too.
• SCAPI tool descriptions are shorter; code-mode details (spec objects, execute globals, examples) moved into the SCAPI skill. scapi_execute is described as the general path for Admin/Shopper tasks when no dedicated tool fits.
• Server instructions and cartridge_deploy point first-deploy prerequisite questions to the config skill.

SDK fix

• SCAPI code mode accepts code ending in a semicolon or a trailing line comment (previously a syntax error).

Latest results (Sonnet 5.5): compliance cases 1.00; enhancement mean Δ +0.61.

Testing

Requires the claude CLI with plugin eval.

1. pnpm --filter @salesforce/b2c-tooling-sdk run build && pnpm --filter @salesforce/b2c-dx-mcp run build
2. pnpm run mcp:eval --runs 1 -j 5 — compliance cases should pass.
3. pnpm run mcp:eval:ablation --runs 1 -j 5 — each case should show a positive Δ.
4. In an MCP client, ask for a SCAPI task (e.g. take a product offline) and confirm the agent reads skill://mcp/scapi/SKILL.md before scapi_search.