[HanielUlises]

A policy fixes an order over its actions. Where two consecutive nodes had no reason to be ordered, the mission waits for nothing, and until now it waited anyway: the tree is a chain of sequences and the second action starts when the first has finished.

parallel_groups finds the runs where that costs something — maximal chains of single continuations whose actions are pairwise independent in two senses:

• classically, answered by the builder from the domain: neither action's effects touch a predicate the other's requirements or effects mention;
• epistemically, answered in the pass itself and conservatively: the agents the two actions name must be disjoint, and neither node's knowledge requirements may mention an agent the other names.

Each run renders as one Parallel, with what followed its last member following the group. Members keep their own CheckKnowledge and their own ApplyEpistemicUpdate, so a run that was not as independent as it was taken to be fails a guard rather than running past it, and the executor answers that by replanning. The second test is exact for ontic, private and semi-private actions and not for a public announcement, which reaches agents it does not name; that case is what the guards are left to catch.

Off unless asked for, with epistemic_bt_builder's parallel_dispatch. With it off the rendering is the one this builder produced before the pass existed, which a test pins byte for byte.

The mission that shows it

eplansys_demo's two-site survey: the site survey run twice in different hands, with each scout carrying the instrument its own site reads so that neither half can be done by the other's robot. Six goal conjuncts, four designated worlds, an eleven-node branching policy.

The pass finds three runs, and says so:

dispatching together: (goto_north north), (goto_south south)
dispatching together: (relay_north north relay), (scan_south south)

Measured against Aletheia 0.2.0, north dirty and south clean:

dispatch	mission
a node at a time	24.0 s
independent runs together	17.0 s

Seventeen is not nine, which is what the two halves would take if each ran from the start. The scan of the north site is where the policy branches, and nothing below a branch begins before the branch resolves, so the south half waits for a scan it has no stake in. Lifting an independent action across a branch point is what this does not do.

Checked

• 10 new unit tests over the grouping and the rendering, including the switch-arity regression a group inside a branch caused before the group and its continuation were made one element
• plansys2_epistemic_executor 157 tests, plansys2_epistemic_bt_builder 37, plansys2_epistemic_planner 173, all passing with the linters
• both dispatches run end to end on the two-site mission, with the goal verified against the state the mission left behind