# Capability catalog proof

Head: 3d82dad76693ecb02b2f4404a8acbfb5151b8607
Base: 5307c233b1f0ae72b8894dd39d94019aaf9e7cb8

Before: the three retained behavioral regressions fail against main production code.
After: 68 tests pass in five focused suites. Targeted lint, contracts/server typechecks and formatting pass.

The real HTTP MCP transport was started on an ephemeral loopback port with synthetic provider, settings and parent fixtures. No production state or real provider account was used. initialize and tools/call crossed the actual transport/schema/service path.

Observed responses:
- Default favorites: primary provider lists caller gpt-5 and favorite; alternate instance lists none. omittedModels is 1 for each provider.
- models=all: primary lists gpt-5, favorite and other; alternate lists favorite.
- Provider filter: only the selected alternate provider is returned.
- Empty and null favorites: both return all visible models.
- Hidden models remain excluded in every case.
- Compact select options retain values, default low and current high; boolean false is preserved; redundant labels are absent.

Independent whole-PR reviews: claude-opus-5-5 and gpt-6-sol, round 1, no confirmed blockers.

This is server/MCP behavior with no visual UI change. Full application/provider sessions, remote/tunnel deployment, macOS, Windows, Electron, iOS and Android were not exercised.