feat: capability extensions — evaluatePermissions, MCP verdict tools, pagination & rerank docs

[MyPrototypeWhat]

Summary

Wave 4 (final wave) of the hardening plan:

• evaluatePermissions (core, minor): the strict-deny gate is now a programmable public API — evaluatePermissions(rights, required, ctx?, opts?) with PermissionKey/EvaluateOptions; the four evaluateUse intents are presets over it. Behavior-identical, proven by the untouched existing test suite; historical verdict wording preserved via a preset label and pinned by new tests.
• MCP verdict tools (mcp, minor): stateless evaluate_use and build_attribution — inputs carry the rights fields, so agents can re-ask "may I use this / how do I credit it" for any prior result without server-side sessions. Plus unified REFKIT_<PROVIDER>_KEY env names (all legacy names still honored, tested both ways).
• Docs: controls.page provider-local pagination semantics + dedupe-by-canonicalUrl "load more" recipe (JSDoc + README); docs/examples/semantic-rerank.md runnable BYO-embedding reranker cookbook.

Test Plan

• pnpm -r typecheck — 20/20 clean
• pnpm test:run — 305/305 (evaluate-use suite unchanged = behavior oracle; +13 new tests incl. wording pins, MCP tool tests, env-alias tests)
• Review-verified: placeholder rights fields never read by the verdict path; all 9 env alias pairs correct; cookbook types match core's actual Reranker contract