test(v1): trim to #2086's coverage — the multi-agent e2e set, no unit shadow

The offline additions (episode durability, engine pins, retry semantics,
sealing, verdict validation) come out: coverage parity with the competing
branch is the review's bar, and the surface is exercised live — the kept
additions are exactly its four e2e equivalents (best-of-n, agentic-judge,
multi-agent env, env-server), on top of the untouched base suite. 64 offline
tests, matching #2086.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Author: hallerite