Support Linux development and stabilize playerbot fixtures

[adrunkhuman]

• Run the shared PowerShell tooling on Windows and Linux. Linux bootstrap uses a locally built client and preserves asset hash verification; Docker and artifact-path errors are handled explicitly.
• Isolate scenario settings, inventory, and budgets. Synchronize death and depot checks with observable milestones, and keep provisioning out of persistence restarts.
• Correct depot proximity ranking, route-adjusted hunt scores, critical-healing priority before Oracle departure, and hunt-completion reasons carried through depot work. Add focused regressions for these contracts and the test harness.
• Update existing setup/testing docs and AGENTS.md; no client source changes or new documentation files in the final diff.

Known failures deferred from this migration: magic_training_post_hunt and magic_training_post_hunt_no_overflow. Both expect immediate post-hunt arbitration, while the controller visits the depot first; that detour also changes mana overflow. Resolving that gameplay policy is separate work.

[adrunkhuman]

Full rebuilt Linux gameplay run: 85/87 passed, 2 assertion failures, no timeouts or skips. The only failures were magic_training_post_hunt and magic_training_post_hunt_no_overflow, deferred as described above. The local client source change is excluded.