ci: isolate dataplane conformance from control plane

[lucarlig]

Summary

Run make conformance and make conformance-bless through the standalone external lane: Redis, dataplane, nginx, and the official fixture, without control-plane startup or publisher-schema timing. The workflow installs cf-integration 0.3.1.

The companion harness change discovers all fixture routes and preserves the real tool schemas instead of maintaining an incomplete hardcoded catalog. Regenerating the baseline removes four false findings: the extra-parameter warning and three capability checks that previously hit missing tool routes. The external server baseline now contains 35 findings, down from 39; fixture-direct and client baselines are unchanged.

Document the limits of the baseline in the testing wiki. Strict matching detects regressions; it does not establish full conformance. Catalog-dependent scenarios cannot exercise checks after the dataplane rejects tools/list, and the pinned fixture lacks a custom-header-annotated tool. Those coverage gaps remain explicit alongside existing protocol failures.

Dependency

Depends on contextforge-org/contextforge-dev-tools#13. Keep this PR draft until cf-integration 0.3.1 is published with the fixture discovery fix.

Validation

• make conformance-bless, followed by clean strict make conformance, using the companion PR's local binary and an image built from this PR's dataplane code.
• All 84 scenarios completed: 40 fixture-direct server, 40 external server, and 4 scoped external client scenarios. Strict result: 44 passed, 40 expected failures, no unexpected failures or baseline drift.
• No missing tool or prompt route errors remain in the external report.
• Repository commit hooks passed: format, clippy, dependency/license audit, nextest, and workspace build.
• mdbook build _context/wiki and mdbook test _context/wiki passed.