docs: simplify predator-first turn order (swap inside step, no loop changes)

docs: spec for multi-agent memory scenarios + predator/agent visual & turn-order refresh

docs: refresh engine README for proteus/game/engine relocation

docs: update HANDOFF architecture + paths for game/web layout

test: repoint vendored-engine guard to proteus/game/engine

refactor(cli): split cli.py god-file into cli/ package (commands/ + parser)

refactor: restructure proteus into game/web subpackages

chore(tools): add import codemod for game/web restructure

docs(spec): preserve agent/provider/registry candidates from untracked game/

docs(spec): preserve arena/leaderboard/db design from untracked interface/

docs(plan): proteus game/web restructure implementation plan

docs(spec): proteus game/web restructure design

feat(web): policy memory modes + replay-only panel (drop text) + graceful provider errors

feat(memory): scripted memory policies (survival_dynamic/refuge/food_rush)

feat(memory): generate 100-turn handover memory (was 8)

feat(memory): store + replay food_cells (paint under agents)

feat(pack_evade): observational food cells (near focal + behind predator) + wall/food prompt

docs: spec for food resources + scripted memory policies + LLM-spectate fix

test: align pack_evade memory frame expectation with wall terrain

feat(web): step-through memory replay grid in the memory panel

feat(web): ship per-turn replay frames in the memory payload

feat(memory): memory_frames() reconstructs per-turn color grids

feat(pack_evade): describe walls in observation + expose wall_rects()

feat(pack_evade): wall-aware predator/optimal/persona BFS + '#' legend

feat(pack_evade): seeded wall blocks + focal-only narrow channels

feat(memory): store wall_rects on checkpoint + Scenario.wall_rects() hook

docs: implementation plan for memory replay + pack_evade walls

docs: spec for memory-movement replay + pack_evade wall terrain

chore: gitignore .worktrees/ for isolated feature workspaces

feat(web): provider toggle (from /config) + model-id field replacing the free-text model spec

feat(web): show the handover memory block in the UI at the start (expandable panel)

feat(runtime): auto-regressive play — handover memory every turn + the model's own prior moves in each observation (Markovian -> auto-regressive)

feat(web): selectable handover memory (default/none/persona/generate/latest) + action quota 100

fix(web): spectate playback speed control + auto-fit grid (64x64 centered, small grids unchanged)

docs(pack_evade): record real-LLM smoke (ollama gpt-oss:120b-cloud — run + spectate, passed)

Merge branch 'web-spectate'

feat(pack_evade): persona weight-vector memory generation (drop manual; reference policy self-play per agentness design)

docs(pack_evade): handoff — 64x64 multi-cell evasion scenario shipped

feat(pack_evade): hand-authored handover memory + default_memory wiring

docs(spectate): handoff — watch an LLM play in color shipped

test(spectate): offline import-safety still holds (providers lazy)

feat(spectate): frontend — mode toggle + live analysis + auto-play/pause/step

feat(scenario): render_frame hook — compact pack_evade obs, predator_evade unchanged

feat(engine): footprint-aware focal bounds (1x1 unchanged)

feat(spectate): /spectate routes + provider config (lazy-imported)

test(pack_evade): pin optimal/chase/reward/no-diagnostic behaviour

feat(pack_evade): 64x64 multi-cell scenario — build, footprint eat, center-Manhattan

test(spectate): golden — SpectateSession trace == SessionRunner(VanillaAgent)

feat(spectate): SpectateSession — agent-driven stepwise LLM driver

docs(plan): web LLM spectating — 7-task TDD implementation plan