[dominikhorn93]

Summary

A splice into a hand-made drawing pushed every shape right of the insertion line, like the modeler's space tool. Rows that neither follow the new node nor stand in its way moved too: a task added after "Abweisung melden" on a rejection path moved the parallel join, "Einlagerung gemeldet" and "Ware eingelagert" on the main path above it by 150 px. Now only the end event of the rejection path moves.

The local push

• PlacementObstacles.pushSet decides what moves. That is what follows the new node beyond the line: along the drawn flows and boundary paths, with boundary events, sub-process content, associated annotations and data, and the groups around them. It also takes, in turn, what a pushed shape or the new node would come close to, sweep over (a line it would jump over) or put on a line. SpaceTool takes this set as push. Beyond the line only listed shapes move, and only connections touching a changed shape bend.
• SpaceTool.capGrowth: pools and lanes, with or without a pool, grow only as far as their content now reaches and keep their margin.

What the narrower push needs, because other rows no longer move along

• A new flow that would enter a drawn node left of its source (no loop) moves that node and what follows it right of the source (NeighbourPlacer.pushBehind). This applies to connect and to every placement in a batch. A merge whose successor lies left of a predecessor is pulled when it led forward into it.
• A batch measures the gap and row spacing the way the same ops do as single commands: flows that a later op re-ends count at their old ends (NodePlacer.drawnEnds). The old global push had hidden a 4 px batch/single difference here.

Fixes for what the narrower push exposed. Each was built and reviewed on its own and measured on the edit benchmark.

• CrossingRepair no longer lays a hand-drawn line along other drawn lines to save a crossing.
• A kept association that a new node or its label touches is routed again (EdgeRouter.touchesFresh; also after LabelRefresh). An association docks on the cheapest segment of its flow.
• LabelPlacer takes a spot that is clear for the one-line text bpmn-js draws before it falls back to the wrapped box.
• RowChooser: a trial route through a label costs 100 instead of 60 (LABEL_HIT).

Two further attempts at the row-opening cost were dropped: one broke batch parity, the other made other runs worse.

Edit benchmark (764 runs, against main at b998c7d): 70 runs better, 11 worse. Runs adding crossings 105 -> 87, median share of moved nodes 30 % -> 26 %, E1 (insert a task): runs adding crossings 11 -> 2. The soft budgets hold. Of the 11 worse runs:

• 8 add nothing to the input. Main resolves problems already in the input by moving unrelated shapes and routing their lines again (artifacts__01 E1/E7, controlflow__s19 E8/E9, subproc__s19-lanes-loop-3 E1/E7, agent__anno1 E9, agent__o2c.e3 E9).
• agent__claim.nodata / .plain E7: "Send rejection" now stays, and a boundary event's label touches it.
• artifacts__16-annotations-events-loop E9: one crossing more than main (20 vs 15).

The docs (docs/layout.md), the CLI guide (AgentGuide.ts) and the plugin skill describe how room is made now.

Checklist

• npm run typecheck passes
• npm run gate passes (2,389 tests, layout regression 370 of 370, fuzz 0 errors / 0 warnings, iso within the budget of main)
• src/report/guide.ts and the matching page in docs/ are updated (only if a command, option or output changed)
• PR title is a changelog-worthy Conventional-Commit summary

Not run: the live-engine tests (no Camunda engine running).

Linked issue

None.

🤖 Generated with Claude Code