Route every document mutation through a commit() helper

[kmatzen]

Completes #54. Stacked on #58 — that PR fixed the instance, this removes the bug class. Base is fix-toggle-node-history, so as with #59 and #60 there will be no CI until #58 merges and GitHub retargets this.

The problem

The history ritual — slice off redo entries, push a deep clone, advance the index — was open-coded at seventeen sites:

const newHistory = state.history.slice(0, state.historyIndex + 1);
newHistory.push(cloneTree(newTree));
set({ tree: newTree, history: newHistory, historyIndex: newHistory.length - 1 });

toggleNode simply omitted it. That's the bug #58 fixed: tree and history[historyIndex] diverged, so an undo discarded one more edit than the user asked for. Seventeen copies of the same four lines is an invitation to write an eighteenth that forgets.

The change

function commit(state, tree, extra = {}): Partial<ModelerState> {
  const history = state.history.slice(0, state.historyIndex + 1);
  history.push(tree ? cloneTree(tree) : null);
  return { ...extra, tree, history, historyIndex: history.length - 1 };
}

Every mutating action becomes set(commit(get(), newTree, { ...view state })). extra carries selection and expansion, so no action touches history or historyIndex directly — which is exactly how toggleNode came to skip them.

addNodeFromData already had a local commit() doing this plus its own expansion bookkeeping — the duplication had been noticed once and solved in one place. I renamed it place() and had it delegate to the module helper; keeping the name would have shadowed it and silently left that action on the old path.

Verification

check	result
tsc --noEmit	exit 0
store suite	28/28 pass
newHistory references remaining	0

The 23 pre-existing tests are the real check here. Behaviour should be identical, and they are what would catch drift. The refactor was applied by a script performing 25 exact-anchor replacements, each asserted present before substitution — so a missed site fails loudly rather than silently. But anchors only prove I edited what I intended, not that semantics are preserved; the tests are what establish that.

No discrimination check applies — this is a pure refactor with no behaviour change to detect.

One sharp edge

commit spreads extra before the history fields, so a caller passing tree or history inside extra would have it silently overridden. The ordering is deliberate — history bookkeeping should win — and no current caller does it, but it is worth knowing before someone tries.

🤖 Generated with Claude Code

[github-actions]

🚀 Preview deployed: https://commit-helper-refactor.sinter.pages.dev

(updates on every push to this PR)