v2.40.0 — world-model bootstrapping + locate_code

Two new default-off, env-gated faculties:

Per-agent world models (adk/worldmodel.py): every agent can grow its own behavioral model from its own experience — loads its checkpoint on construction, records one transition per tool per turn, promotes itself cold -> warm -> trained. Modes: off (default) / learn / shadow / steer; shadow is verified byte-identical to off, and steer only ever applies a strict permutation. Pure stdlib; register a custom backend via register_world_model(). CLI: adk wm status/inspect/train/reset. Docs: docs/WORLD_MODEL.md.

locate_code (adk/code_locator.py): one indexed lookup instead of a grep chain. Ensemble of structural + semantic code search with cross-lane agreement ranking; measured 83% localization-token saving on a large monorepo (~150 tokens per hit vs ~11k per grep chain). Enable with AITHER_CODE_LOCATOR=1 and point AITHER_CODEGRAPH_URL at your index.

Both degrade to no-ops when unconfigured. 81 new tests.