feat(errand): handover memory carries courier position + behaviour flags on the text path

feat(errand): no move limit — ends only on reaching the house (analysis) or zero health

feat(errand): surface grass-cut/avoid + pedestrian-touch in results; grass breaks civic/outlaw persona tie

feat(errand): director grass-aware persona routing — civic avoids lawn, cut personas take the grass shortcut

feat(errand): results summary (event choices, closest persona, headline metrics) in review

test(errand): memory renders walls, donut, single courier

feat(errand): single-agent BFS-routed memory + 3 persona variants on MEMORY_LAYOUT

test(discovery): end-to-end errand_runner session emits discovery metric

feat(memory): author_errand_runner persona demo (cells overlay + npc pedestrian) + default_memory

feat(discovery): discovery_turn/identified/efficiency metric (additive)

feat(discovery): source available actions from scenario.action_set (interact reaches the agent)

feat(discovery): parse SELF: report + score self_correct in make_turn_trace

feat(discovery): TurnTrace self_belief/self_correct + Scenario discovery hooks

feat(memory): per-turn cells overlay channel + npc_down/npc_active colours in memory_frames

feat(director): predator_chase pack scatters (all survivors flee) after the first kill

feat(director): predator_chase memory — longer pre-kill preroll + wider, varied agent roaming

fix(director): agents never overlap in authored memory (agent-agent collision avoidance)

test(template): restore generic eliminated-outcome + blocked step_reward coverage

refactor(scenario): delete predator_evade; template is the canonical scenario

refactor(scenario): rename pack_flee -> predator_chase

refactor(scenario): rename pack_evade -> template

fix(resource_race): collected resource leaves the field (live sprite removal + memory pickup frame)

fix(memory): legacy replay block sizes match resized pack_evade (3x3 predator, 2x2 focal)

feat(director): author_resource_race multi-agent winner memory

feat(director): author_pack_flee multi-agent survivor memory

feat(memory): multi-agent memory_frames (2x2 agents, 3x3 mouth predator, resources)

feat(memory): AgentFrame model + multi-agent turn/checkpoint fields

refactor: restructure proteus into game/web subpackages

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

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

test: align pack_evade memory frame expectation with wall terrain

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

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

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

Merge branch 'web-spectate'

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

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

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

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

Merge branch 'master' into web-interactive-play

feat(cp8): persona-driven memory generation (hidden weights, public id)

feat(cp8): persona-maintenance metrics (agreement/regret/pressure-weighted/drift)

feat(cp8): hidden persona weights + R_w reference policy + pressure

feat(cp8): record turn_order/capture_rule/horizon on the trace

feat(cp8): time_to_capture/distance_auc/min_distance/near_capture_count metrics

feat(cp8): record pre/post BFS distance + post positions per turn

feat(cp8): additive per-turn distance/post-position trace fields

test(web): golden — InteractiveSession trace == SessionRunner(HumanAgent)

feat(web): InteractiveSession — threadless stepwise play driver

feat(cp7): hybrid memory injection into SessionRunner turn-1 observation