feat: Phase F3 declared snapshot/restore capability with measured-fork branching and honest refusal

docs/NEXT_STATUS.md

@@ -33,7 +33,7 @@ language — this table is the current status source of truth.
 | C | partial | Dossier + cited gate (#37), orphan-label refusal + provenance + fault robustness (#42) | C4 world-native distributions (need SOFA/GPU world revisions); C5 phantom (blocked: partner) |
 | D | partial | Paired comparison + CIs (#43), compare CLI (#45), split manifests with patient/site grouping (#60-62), D5 explicit cohort subgroup validation, head-covered trial bindings, clustered continuous bootstrap, and cloud parity (#64, cloud #5) | D5 benchmark-level reporting with external study data |
 | E | partial | Obs/action contract + A3 (#44), interactive streams (#46), LeRobot reader + MONAI (#51), honest video adapter (#54), semantic domain profiles for StreamSpec and CapabilitySpec (#66), profile-binding soundness closure: channel/schema conflation, identity pinning, geometry and calibration validation (#67) | E4 media alignment (no test media; mp4 banned from VC); real robotics policy adapter |
-| F | partial | Prefix-replay branching proof + seed caveat (#47), trajectory-backed forecast task + null baseline (#56), planning utility measurement and recipe (#58) | F1 trajectory benchmark imports; F4 closed-loop simulator branching |
+| F | partial | Prefix-replay branching proof + seed caveat (#47), trajectory-backed forecast task + null baseline (#56), planning utility measurement and recipe (#58), declared snapshot/restore capability with measured-fork proof and honest branch-equivalence refusal (#68) | F1 trajectory benchmark imports; F4 closed-loop simulator branching |
 | G | partial | SB3 PPO recipe + measurement (#48), runnable fixes (#50, #52), kernel-observed episode divergence (#57, #59), verifier-derived training rewards with divergence penalty (#65) | Prime interop |
 | H | partial | Episode resume (#49), crash-safe writes + partial recovery (#53), strict trial provenance + pre-write drift rejection (#59) | Fleet queue/autoscaling (blocked: hosted capacity decision); vectorized stepping (not justified by measured need) |
 

src/or_audit/eval/__init__.py

@@ -15,6 +15,12 @@
 )
 from or_audit.eval.agent import AgentPackage
 from or_audit.eval.bind import assert_bind
+from or_audit.eval.branching import (
+    BranchEvidence,
+    branch_from,
+    measure_branch_support,
+    require_branch_support,
+)
 from or_audit.eval.cartesian import CartesianManifest, replay_cartesian, run_cartesian_job
 from or_audit.eval.contracts import (
     CapabilitySpec,
@@ -98,9 +104,11 @@
 from or_audit.eval.vector import TrialVector, project
 from or_audit.eval.worlds import (
     WORLD_KIND_ENTRY_POINT_GROUP,
+    BranchingSupport,
     DeterminismClass,
     WorldCapabilities,
     WorldKindSpec,
+    branching_at_least,
     determinism_at_least,
     list_world_kinds,
     require_world_kind,
@@ -117,6 +125,8 @@
     "AttestationLevel",
     "BaseModalityAdapter",
     "BaseSimulationBridge",
+    "BranchEvidence",
+    "BranchingSupport",
     "CapabilitySpec",
     "CartesianManifest",
     "DatasetSpec",
@@ -172,6 +182,8 @@
     "assemble_job_result",
     "assert_bind",
     "bootstrap_mean_ci",
+    "branch_from",
+    "branching_at_least",
     "builtin_random_agent",
     "clear_adapter_registry",
     "clear_simulation_registry",
@@ -193,6 +205,7 @@
     "make_isaac_bridge",
     "make_sofa_bridge",
     "make_warp_bridge",
+    "measure_branch_support",
     "preprocess_observation",
     "project",
     "reconstitute_trial_vector",
@@ -202,6 +215,7 @@
     "replay_cartesian",
     "replay_job",
     "require_adapter",
+    "require_branch_support",
     "require_simulation_engine",
     "require_world_kind",
     "reset_default_simulation_engines",