feat(runtime): execute a plain flow as a streaming transfer

[devin-ai-integration]

What and why

A flow between two action parameters ran as a succession flow: applyDataFlows moved the source pin's value to the target pin when the source node completed, for both spellings. The parser kept the distinction (ast.Usage.IsSuccessionFlow) but lower.ObjectFlow carried no kind, so it was lost at lowering. A consumer that reads its pin before its producer completes got the wrong result.

• lower.ObjectFlow.Kind (FlowStreaming, the zero value, or FlowSuccession) is set by lowerFlow from the declaration; succeedFlow reads the kind rather than the declaration to add the succession edge a succession flow also states. ast.ObjectFlowEdge (never a succession) lowers as streaming. The runtime never looks at Decl.
• Streaming (runtime/action_frame.go): setFrameFeature → streamFrom → streamFlow carries each write to a source pin along the node's streaming flows, to the pin of every ongoing performance of the target (ongoing, the performances of a node under way). With no performance under way the value goes through deliverFlow and waits at the pin as before, one per performance, oldest first — which is what makes a target that has not begun, a stream inside a loop body and fork/join around the two nodes work. actionFrame.streamed records the pins streamed so applyDataFlows does not carry the last value again at completion.
• Succession: unchanged — applyDataFlows → deliverFlow at completion, and the target begins after the source.
• Typed errors: a source that completes with the pin never written is ErrFlowSource (wrapping the message the flow-source case already reported); a write after the target's last performance ended, which no later performance of the target takes, is ErrStreamUnreceived when the enclosing performance completes (actionFrame.unreceived, cleared by takeDeliveries, checked in endPerformance and at root completion); a stream to a pin the ongoing target does not declare is ErrNodePin at the write. streamed/unreceived are carried through snapshots, held images and check state spelling.
• Shared executor plumbing: only applyDataFlows gains a streamed parameter (three call sites: leaveExecutionNode, completeNode, the body-node path in action_statements.go). Nothing in state_executor.go or the inline do bodies changes.

Known limitation (recorded in the roadmap): an atomic body is performed whole in one step, so a source that writes its pin several times before any performance of the target is under way queues every value and each target performance takes one; a truly concurrent target would see the last.

Specification basis

SysML v2 (formal/2026-03-02) §7.16: "the input and output parameters are streaming unless designated as succession flows"; a streaming flow "can be ongoing while both the source and target action are being performed", a succession flow "cannot begin until the source completes". Library: Flows::Flow :> Message, FlowTransfer vs Flows::SuccessionFlow :> Flow, FlowTransferBefore. The Actions map's object-flow row in docs/project/spec-compliance.md is split into a streaming row and a succession row (both ✅), "Streaming pins" leaves "Major Features Not Implemented", and the roadmap's streaming-flows item is recorded as landed.

How it was verified

• lower/action_flow_kind_test.go: the kind follows the spelling; only the succession kind adds an edge.
• Conformance: action_flow_streaming_producer_consumer (producer loop writing 1, 2, 3 beside a consumer loop: total 6) + trace golden showing each write landing before the next read; action_flow_succession_producer_consumer (same nodes in sequence: consumer reads 3 three times, total 9) + trace golden; action_flow_streaming_before_target_begins; action_flow_streaming_in_loop_body; action_flow_streaming_declared_value (a declared out value = 5 streams as the producer begins).
• robustness_streaming_flow_test.go TestRuntimeRobustnessStreamingFlow: source_never_writes (ErrFlowSource), target_completed_before_source_writes (ErrStreamUnreceived), target_pin_not_declared (ErrNodePin), later_performance_takes_one_of_two_late_values (the second late value is still reported after one performance takes the first).
• modelform.TestGraphsActionCarriesTheFlowKind: the graphs:1 form's ObjectFlowForm.kind spells both kinds (added field, version unchanged).
• go build ./..., go vet ./..., gofmt -l . empty, go test -count=1 ./... green with OPENSYSML_REQUIRE_TRAINING_CORPUS=1 OPENSYSML_REQUIRE_PILOT_CORPORA=1 and the corpora downloaded (no ratchet movement; training_examples_expected.txt untouched), TestPilotLibraryXMI with its XMI downloaded, python3 scripts/changelog.py check, scripts/check-doc-ids.py, doc-counts -check.

Checklist

• make test and make lint pass locally
• Tests added or updated for the change
• Documentation extended where it already covers the surface (see CONTRIBUTING.md)
• Changelog entry added as changes/unreleased/<slug>.<section>.md, not as an edit to CHANGELOG.md
• baselines regenerated and make docs-counts run if a gate count moved (compliance rows need nothing: the census is counted at docs build)
• No internal work-item labels (waves, slices, F4, K5) in the body, docs, or changelog

[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