Fix cache-corrupted token metrics in multi-turn Claude Code eval

[prernakakkar-google]

Problem
Multi-turn claude_code evaluations drove each turn with a fresh claude -p --resume --fork-session subprocess. Prompt-cache continuity does not survive across separate claude -p invocations, so every turn after the first re-paid a cold cache_creation cost instead of reading from cache. This inflated cache_creation_input_tokens and corrupted the token metrics the eval exists to measure (tokens_processed, effective_billed_tokens). (known issue)

Fix
Drive an entire scenario through one persistent claude process using Claude Code's streaming-input mode (-p --input-format stream-json --output-format stream-json --verbose): each turn is a newline-delimited user message on stdin, and the process stays alive across turns, so session context and the prompt cache persist. Closing stdin ends the session.

Introduces a small per-scenario session abstraction (AgentCliSession) so the change is backward-compatible:

1. Default _PerTurnSession reproduces the old --resume-per-turn behavior exactly → gemini_cli / codex_cli / agy_cli are unchanged.
2. ClaudeCodeGenerator overrides start_session() with a persistent streaming session (background stderr drain to avoid pipe deadlock, per-turn watchdog timeout, clean stdin-EOF shutdown, error path when the process dies before a result). Reuses the existing _parse_stream_json unchanged.

Also included (related: Vertex cache accuracy)
Default Claude Vertex calls to the global endpoint (vertex_region: "global"); regional pinning worsens cache-hit determinism and is unsupported past Sonnet 4.6.

https://paste.googleplex.com/newui/?id=5943794385158144

[prernakakkar-google]

/gcbrun