test(e2e): pass mock LLM handlers to startModelRecorder

[danielle-mustillo]

Why

The test-e2e (microvm) job fails at random on egress tests with unexpected EOF (job, job). The egress and mounted-skills mocks put a second server with an unbuffered httputil.ReverseProxy behind startModelRecorder. Once the streamed response starts, that server closes the request body. The proxy transport then reads the body once more past Content-Length, gets ErrBodyReadAfterClose, and closes the upstream connection mid-stream.

What changed

• startModelRecorder takes an http.Handler and calls it in process with the buffered body, so no proxy reads a body that the server can close.
• The egress and mounted-skills mocks pass their routing handler directly and lose their second server. Other callers wrap the mock URL with mockLLMProxy.
• TestModelRecorderBodyOutlivesResponseStart checks that the upstream handler can read the body after the response starts.

Scope

• Fixes the use of closed network connection and unexpected EOF flake only.
• Does not fix the separate live fixed did not become ALLOWED flake in the claude variant.

Tradeoffs

• Calling EnableFullDuplex in each mock also fixes the race, but each new mock has to remember it. Removing the second hop removes the bug class.

Blast Radius

Test code only. Every e2e test that uses startModelRecorder changes, but the request path is the same minus one local hop.

Verification

• TestModelRecorderBodyOutlivesResponseStart passes, 50 of 50 runs. It fails with http: invalid Read on closed Body if the recorder forwards the unbuffered body.
• A local repro of the old chain with a 50 ms delay before body EOF truncated 200 of 200 responses.
• test-e2e (microvm) on this PR passed 3 of 3 attempts.