[RL][Phase 5] Implement the fixed-deck mirror-match episode runner

[utilForever]

Parent: #1057
Epic: #1052

Goal

Run one complete self-play episode using the same fixed, low-randomness deck for both players.

Scope

• Configure the agreed MVP deck fixture for both seats
• Use the fixed non-learned mulligan policy
• Invoke PUCT/ISMCTS for MAIN_ACTION decisions
• Record one pending sample per decision and attach the final outcome
• Stop cleanly on win, loss, draw, or configured safety limit

Non-goals

• Arbitrary deck matchmaking
• Learned mulligan policy
• Parallel episode workers

Acceptance criteria

• An episode reaches a valid terminal result without illegal actions.
• Both seats use the same configured deck fixture.
• Each decision produces one valid training sample.
• Fixed game and search seeds reproduce the episode.

Verification

Add one deterministic C++ episode smoke test with the uniform evaluator.