fix(binding-mcp-openapi): resolve exit per route instead of one binding-wide value

[jfallows]

Description

Repurposed from the original "lock down the schema" approach — see discussion below for why.

Follow-up to #2563. That PR made the kind: proxy exit requirement internally consistent, but the underlying implementation still only ever resolved a single binding-wide exit, silently discarding any other value declared on a route (findFirst() over all routes, applied uniformly to every generated mcp-http0 route via McpOpenapiCompositeGenerator.injectRoutes()).

The original fix for this (commit dee960ff, since reverted here) tightened the schema to forbid routes[].exit entirely and require exit only at the binding level — mirroring binding-mcp-kafka-connect/binding-mcp-schema-registry. That approach was rejected: it's a backward-incompatible schema change — any existing config using route-level exit (a shape #2563 explicitly allows and requires in the no-binding-level-exit case) would fail validation outright. Comparing against binding-mcp-kafka, which genuinely supports different routes exiting to different downstream bindings (some tools bypass its local cache and go to a live Kafka client, others don't — a real structural need, not just permissiveness), it's more correct to fix mcp-openapi's implementation to actually deliver on what its schema already promises, rather than restricting the schema to match a buggy implementation.

This PR instead implements genuine per-route exit resolution, keeping the schema exactly as merged in #2563:

• Fixed a live regression along the way: McpOpenapiBindingConfig.routes filtered on route.exit == null to exclude the engine-synthesized catch-all route (for the top-level exit: shorthand). But since real routes can also carry their own exit (per #2563's schema), that filter also excluded any real route with its own exit — silently dropping its with/when from the generated composite. A proxy config using only per-route exits (no binding-level exit) currently produces an empty mcp-http0 binding with zero tools/resources. Fixed by identifying the synthetic route via with == null instead (every real mcp-openapi route always has with; the synthetic one never does).
• McpOpenapiRouteConfig now carries its own resolved exit.
• McpOpenapiCompositeGenerator resolves each RoutedOperation's exit individually — the originating route's own exit, falling back to binding.exit — and applies it per generated route, instead of broadcasting one binding-wide value onto every route regardless of origin.

Test plan

Test-first: added shouldGenerateProxyKindMcpHttpFromRouteExit first and confirmed it failed against the unfixed code (mcpHttp.routes came back empty — proving the live regression), then implemented the fix.

• ./mvnw clean verify -pl runtime/binding-mcp-openapi: 61/61 unit tests (including three new cases: a route with its own exit is no longer dropped; two routes with different exits produce two generated routes with correspondingly different exits; a route with no exit falls back to the binding-level one) + 26/26 ITs (McpOpenapiProxyIT, McpOpenapiClientIT) pass
• ./mvnw clean verify -pl specs/binding-mcp-openapi.spec: 14/14 SchemaTest cases (unchanged from fix(binding-mcp-openapi): correct proxy exit schema and route filtering #2563) + 17/17 spec ITs pass
• Checkstyle clean, JaCoCo coverage checks met, NOTICE up to date on both modules

🤖 Generated with Claude Code

https://claude.ai/code/session_01SBDndP49e5oZjtxYNmzH5h