Make upstream mocks prove real streaming and current provider/MCP contracts

[ebursztein]

Outcome

Consolidate #145, #146 and the residual coverage requirements of #150 into one mock-fidelity change. A test upstream must actually exercise the streaming behavior its routes advertise.

Current evidence

Inspected main at 914c72e: capsem-mock-server/src/main.rs still routes /chunked and /delayed-chunks through the same buffered response helper; /sse/model also uses a buffered response. The original #150 claim of only three tests is stale: src/tests.rs now covers static routes, provider JSON/stream terminators, MCP dispatch, token/path parsing and other shapes. Preserve those tests; fill remaining gaps rather than rebuilding them.

Required work

• /chunked emits multiple body frames with genuine HTTP/1.1 chunked transfer, no fixed Content-Length. /delayed-chunks additionally exposes controllable gaps.
• SSE/model streaming routes deliver incremental events rather than one completed buffered blob; preserve each provider's event/tool-call/terminator shape.
• Streaming proof must observe the first frame before completion and exercise split events, cancellation and backpressure. Merely collecting the final body proves none of these.
• Reconcile [triage:tooling-06-missing-tests-mock-http-routing-shapes] Mock server HTTP routing and response shapes are almost entirely untested #150 against existing tests: retain tool-call/no-tool-call branch coverage, MCP errors, fixture target extraction, request-log fields/header normalization, /log body omission and response-byte accounting. Add only missing behavior checks.
• Verify existing MITM/parser tests consume the streaming path; no second mock server implementation.

Acceptance and cost

Use controlled channels, barriers or paused time to demonstrate delayed/incremental delivery; no seconds-long sleeps per case. Reuse existing public-path fixtures for affected streaming behavior. No new endurance lane. Report which original findings were already fixed and which this change resolves.

Original reports #145/#146/#150 remain provenance; closure as superseded is not a claim of implementation. This is test-infrastructure correctness, independent of product proxy extraction #206 and standalone features.