Make probabilistic FSMs the sole autonomous model

[ppericard]

The default launch now runs the mixed probabilistic FSM population. Every autonomous agent has one inherited graph; flat action-weight strategies, their four preset comparisons, and random-removal mode are removed from configuration, CLI, JSON and UI. Explicit scripted demo mode remains an engine regression fixture.

Pierre has accepted FSMs as the sole autonomous model. This simplification preserves the current integrity/crowding equations, attempted-action memory, inheritance and all-eight destination sampling. MODEL describes the current rules with diagrams and worked examples; the development guide maps modules to tests; contributor guidance records visual explanations and the accepted direction. JavaScript checks now discover files automatically. The cleanup also rejects array-shaped control commands and guards the entire zero-tick test lifecycle.

Validation:

• Debug/release headless tests, web/API tests, both Clippy modes, formatting, documentation checks and JavaScript checks pass locally.
• A complete 200-tick seeded run matches the pre-cleanup executable: IDs, positions, graphs, action memory, integrity, event totals and retained failures. A smaller readable reference is retained as a CLI regression.
• Full Linux and Windows checks, Chromium scenarios and dependency audits: https://github.com/ppericard/virtual-life/actions/runs/34556347945
• Independent read-only Spec and Standards reviews found no blockers. The small README panel-description correction is included. The current inspector was also visually checked; visual inspection and Pierre's acceptance remain separate from automated tests.

Compatibility: --bundles, --survival and retired preset IDs are rejected. Historical Git revisions retain those models. No dependencies or additional model mechanisms were introduced. Mutation is tracked in #18; the guarded-destination/outcome-state alternative in #14 remains unapproved.