test(review): canary that a branch's Claude SessionStart hook does not run

T-243 asked for a regression test proving a committed SessionStart hook does not execute during current-branch review or entire review --target. This drives the real claude CLI through the real review command, with an invalid API key so no tokens are spent: the hook fires at startup, before any model request. Entire's own SessionStart hook recording a session is the positive control that the startup phase ran, and a PATH shim proves the agent was launched, so a run that never got there cannot pass.

Fails against main's reviewer (the hook runs in both cases) and passes with this branch. Opt-in via ENTIRE_TEST_REVIEW_REAL_AGENTS=1.