Commit open edit when displaced by clicking another node (#376)

[CarlosNZ]

Closes #376.

What & why

Opening an edit on another node — clicking its ✎ pencil, double-clicking another value, or clicking another key to rename — currently cancels the in-progress edit. But Tab commits then moves. The two "leave this field and go edit elsewhere" gestures behaved oppositely, so typing a value and clicking the next field silently discarded the typed text.

This makes a click-to-another-node switch behave like Tab.

Behaviour (the contract the tests pin)

Outgoing session	Changed	Unchanged	Invalid
value edit	commit + commitEdit + onUpdate, open new	submitEdit+commitEdit (no onUpdate/setData), open new	block: stay open w/ error
collection JSON edit	commit, open new	commits (ref-eq misses structural no-op)	malformed JSON → block
key rename	commitRename + onUpdate, open new	commitRename (no onUpdate), open new	duplicate key → block
object add	cancelAdd, open new	cancelAdd, open new	n/a

• No-op → commit* (consistent with Tab), incl. no-op rename → commitRename (now flagged via buildCommit isNoOp: oldKey === newKey; also aligns Tab-off-an-unchanged-key).
• Object add is the exception — a displace still cancels it (you can't Tab out of a new-key edit). Implemented simply by not registering a commit callback for add.
• Invalid blocks the switch — the editor stays open with its inline error; Esc / ✗ remain the discard path.

How

• The commit logic lives on the outgoing node but open() is called by the incoming one, so the outgoing node registers a commit-on-displace callback in the editing store, alongside today's cancelOp. open(), on a switch, calls it and defers opening the new node into the commit's onCommit (synchronous for editor ops, so it still feels instant) — no re-entrancy into open().
• The callback forwards to the node's live handleEdit via a ref — a closure frozen at open() time would commit the stale initial buffer (the ref hooks sit above the !isVisible early returns to keep hook order stable).
• Invalid-blocks and no-op fall out of the existing engine: an invalid edit returns before submit (so onCommit never runs → switch blocked); a no-op routes through submit's isNoOp branch.
• submit() now also abandons gracefully if a displace fires for a session whose node already unmounted (buildCommit would otherwise throw on the vanished path).

Touches: EditingProvider (store), ValueNodeWrapper, CollectionNode, KeyDisplay + useCommon (rename), JsonEditor (buildCommit).

Tests & docs

• Rewrote the two cancel-on-displace tests to the commit contract (confirmed failing first) and added 7 new tests covering changed / unchanged / invalid across value, collection, custom-node, rename, plus the add-cancels regression guard.
• Updated the A8 manual-test matrix, added a changeset, and noted the onEditEvent contract shift in the migration guide (§10) + an end-user README bullet.
• ✅ pnpm test (653 pass), pnpm compile, pnpm lint all green.

Note for review

One behaviour change slightly beyond the strict displace scope, for consistency: a duplicate-key rename on Enter now keeps the editor open with the error instead of closing (matching value/collection invalid handling). And the automated suite can't reproduce the Firefox DnD phantom-dragstart path — a quick Firefox walk of the A8 matrix (incl. a drag right after a commit-on-displace) is worth doing before merge, though the draggable flip logic is untouched.

🤖 Generated with Claude Code

[github-actions]

Bundle size impact

json-edit-react

Format	Base raw	PR raw	Δ raw	Base gzip	PR gzip	Δ gzip
esm	56.71 KB	57.27 KB	🔺 +574 B (+0.99%)	20.30 KB	20.45 KB	🔺 +149 B (+0.72%)
cjs	58.19 KB	58.76 KB	🔺 +587 B (+0.99%)	20.33 KB	20.49 KB	🔺 +162 B (+0.78%)

_{Measured from build/index.{cjs,esm}.js. Gzip at level 9.}