[devin-ai-integration]

What and why

Step 3 of the execution-owned IR migration (docs/internals/execution-ir-design.md, "Migration as a stack of pull requests", row 3): the state executor's own records stop holding *ast.StateNode, *ast.StateRegion, *ast.PseudostateNode and *lower.Transition pointers as keys and hold the graph-local ids step 1 assigns. Stacked on #2137 (feat/ir-program-ids), which must merge first; this PR then retargets to develop.

Records keyed by id now: stateConfiguration (simpleState StateID, regionStates map[RegionID]StateID, defaultComplete map[RegionID]bool), stateAttrs/stateAttrCells/stateAttrIndex, stateStack, history/historyRecord, joinArrived/joinChosen, breakpointNodes map[VertexRef]bool/breakpointHit/pausedAt VertexRef, timerScheduled/timeTriggerVerdict/changeFired/changeObserved/changePending/changeReads/changeRearmed/firingChange by TransID, leftAhead/enteredAhead/entering/activeAtEntry/performing, pendingRouteEntryEffects, doAction.state, dispatchCandidate.leaf/source, route (target, choice, arrival, crossed, origin, terminate VertexRef), junctionDraw.at/origin, routeEffect.within, regionEntry, lazyEntry, unitHead.shared, firingScope, heldEntry, and the mirrors in imagedState (held image) and stateCapture (snapshot). Where a consumer still needs the declaration it goes through Provenance (state_ids.go: stateID/stateDecl, regionID/regionDecl, pseudoID/pseudoDecl, transID/transDecl, vertexRef/vertexDecl). The graph keeps its AST fields; no layering change.

Public surface:

• New program.VertexKind (VertexNone, VertexState, VertexPseudostate, VertexTerminate) and program.VertexRef{Kind, Index} with StateVertex, PseudoVertex, TerminateVertex, IsZero, State, Pseudo, Terminate. A terminate action's index is its 1-based ordinal in StateGraph.Terminates, which lowering does not number yet.
• runtime.FiredTransition is now {Transition TransID; Source, Target VertexRef; Entry EntryRef; Origin program.Origin}; new runtime.EntryRef{Owner StateID; Region RegionID; Index int} names an entry transition (lowering gives those no id yet) by its body and 1-based position. New (*StateExecutor).FiredNodes(FiredTransition) (decl, source, target, owner ast.Node) is the back-map the LSP debugger uses.
• ActiveStates, ActiveLeaves, CurrentState, StateStack, EnclosingStates, StatePath, RegionOf, SetBreakpointAt, PausedAt keep their AST signatures and translate at the boundary (step 7 moves those consumers).

Where the design met the code:

• Hidden states (the graph-only state a parallel body's typed usage stands for) are not in StateGraph.States, so step 1 gave them no id, yet the executor reaches them through ParentState, RegionState, HiddenRegionOf and StateAttributes. assignStateIDs now numbers States first and then those states (statesInOrder), so len(Origins) == len(ID)+1; the step 1 denseness test asserts against ID rather than States. The step 1 API is otherwise unchanged.
• The machine state (StateGraph.Machine) is also outside the state table; the executor gives it the id one past the last numbered state (machineID). Step 8/9 should fold it into the table.
• Entry transitions have no TransID/origin in step 1, so FiredTransition.Origin for one carries only the declaration's doc and span, with Kind: "entry transition".
• Internal helpers that read AST-keyed graph tables keep declaration-typed parameters until step 8 removes those tables; the records they read and write are id-keyed.

No snapshot format bump: the snapshot is an in-memory capture and no serialized form changed. No conformance .expected.json, golden trace or ratchet moved.

One baseline moved: the gzipped wasm size budgets in tests/wasm (replGzipBudget 16100000 → 16200000, engineGzipBudget 9500000 → 9600000). On the base both binaries fit; with the id tables' accessors at every record read the RE REPL build measures 16108375 gzipped bytes and the engine 9526983, so the budgets were raised with headroom for steps 2 and 5, which add the same kind of code.

How it was verified

• go build ./..., go vet ./..., gofmt -l . empty, make lint
• go test ./... (with cmd/sysml, tests/codegen, tests/export run alone with -timeout 30m), go test -count=1 ./tests/corpus -run 'TestTrainingExamples|TestPilotCorpora' after the download scripts, go test ./tests/hygiene, go test -C tools ./... (PSSM referee record over the pinned suite)
• Conformance state cases, TestExecutionTrace under every region-order policy, TestRuntimeRobustnessRegionOrder, …Terminate, …ResumableInlineDoBody unchanged
• New: TestVertexRef* (program), TestAssignIDs_StateHiddenRegionOwnersNumbered (lower)
• Comment check (git diff --merge-base origin/feat/ir-program-ids -- '*.go' | awk ...) prints nothing

Checklist

• make test and make lint pass locally
• Tests added or updated for the change
• Documentation extended where it already covers the surface (none covers these internals)
• Changelog entry: none, internal refactor
• baselines regenerated and make docs-counts run if a gate count moved (none moved)
• No internal work-item labels (waves, slices, F4, K5) in the body, docs, or changelog

Link to Devin session: https://nasa-jpl-demo.devinenterprise.com/sessions/a4a5e131eec042ecb18c40d69306af1a
Open in Devin Desktop: https://nasa-jpl-demo.devinenterprise.com/desktop/session/a4a5e131eec042ecb18c40d69306af1a?variant=devin
Requested by: @HuiJun

---

🟢 Devin fixed the finding · Follow Devin's work

[devin-ai-integration]

I'll fix CI failures and address comments from users with write access. I'll skip comments containing "(aside)".

• Disable automatic comment, CI, and merge conflict monitoring

[devin-ai-integration]

Devin Review found 1 finding. Devin automatically fixed it.

Fixed by Devin (1)

• Nested parallel regions share a zero ID — statesInOrder now reaches hidden owners through a worklist over parent chains and composite regions, and regionsInOrder walks it instead of States, so a hidden owner's regions take ids. Lowering currently rejects a parallel typed usage inside a parallel body ("region a has no initial state"), so no fixture reaches the case; the numbering is complete regardless.

View findings in Devin Review · Follow Devin's work