feat(cp6): CLI compare — aggregate human/LLM traces by (model, difficulty)

feat(cp6): runtime.aggregate — pure (model,difficulty) trace aggregation

feat(cp6): trajectory + category-reward metrics (away_move_fraction, trajectory_agreement, ...)

feat(cp6): runtime.rollout — deterministic optimal-trajectory simulator

feat(cp6): scenario-owned survival step_reward (away-from-predator) + safety_distance

test(cp6): lock optimal!=habit handover invariant for all difficulty bands

feat(cp6): per-difficulty hand-authored predator_evade layouts (EASY unchanged)

feat(cp6): Scenario.build_level takes difficulty + record_focal_move/safety_distance defaults

docs(cp6): next-session execution kickoff prompt (subagent-driven)

docs(cp6): implementation plan — 9 TDD tasks (layouts + category eval + compare)

docs(cp6): design — difficulty layouts + category-specific eval + human-baseline compare

docs(cp5): note stdin-EOF fix + empty-turns terminal-flag deferral (103 passed)

fix(cp5): _cmd_play returns rc=2 on stdin EOF (no traceback, no partial-trace crash)

docs(cp5): handoff close — human play + trace viz shipped; CP6 next

feat(cp5): CLI replay --visual/--png/--fps (text stays default)

feat(cp5): CLI 'play' subcommand — human session via stdin

test(cp5): guard viz package against eager matplotlib import

refactor(cp5): lazy matplotlib in arc_grid.rendering (offline import invariant)

fix(cp5): close PNG figures on error + force Agg + _title unit tests

feat(cp5): viz.png — per-frame titled PNGs via matplotlib Agg

fix(cp5): exercise fps->delay path in terminal test + respect explicit empty color_map

feat(cp5): viz.terminal — truecolor block grid + CP4.5 side panel

fix(cp5): harden viz.reconstruct cut verify + make FrameStep hashable

feat(cp5): viz.reconstruct — deterministic frame replay with double self-verify

test(cp5): lock human↔LLM trace comparability invariant (spec §10)

feat(cp5): HumanAgent probe path returns typed ProbeResult

feat(cp5): HumanAgent act path + WASD parsing + reprompt (I/O-injected)

docs(cp5): next-session execution kickoff prompt (subagent-driven)

docs(cp5): implementation plan — HumanAgent + viz (11 TDD tasks)

docs(cp5): brainstormed design — HumanAgent + trace viz (engine-replay reconstruction)

docs(cp4): handoff update — CLI + cloud providers + token accounting shipped; CP5 kickoff prompt; gitignore .env

feat(cp4.5): preserve probe reasoning + raw_text + token accounting via ProbeResult

feat(cp4.5): persist act-call token accounting + raw_text in TurnTrace

feat(cp4): CLI (run / list-scenarios / replay) over fake + real providers

feat(cp4): JSONL trace persistence (append_trace / read_traces)

feat(cp4): port cloud providers (openai/anthropic/gemini/ollama) into proteus

feat(cp4): provider factory (make_provider) with lazy registry + fake path

docs: CP4 implementation plan (real-provider CLI)

docs(cp3): handoff record + CP4 pointer

test(cp3): golden test asserts full round-trip equality + concrete accuracy anchor

feat(cp3): runtime re-exports + full-session golden integration test

feat(cp3): guard cut-replay termination + cover survived/eliminated/cut-frame contracts

feat(cp3): SessionRunner orchestrator (cut replay -> play -> trace)

test(cp3): cover perfect/all-diagnostic/cap/empty metrics + single-source divergence

feat(cp3): session metrics (motive-reading, reactivity, survival, divergence)

feat(cp3): constrain outcome to Literal, default seed, test nested trace roundtrip

chore(cp3): drop tests/runtime/__init__.py for test-tree consistency

feat(cp3): lean TurnTrace + SessionTrace models

fix(cp2): VanillaAgent parses action from think-stripped answer (ignore in-think decoys)

feat(cp2): VanillaAgent (act + probe, reasoning capture)