[rendis]

Follow-up to the changes merged in #22, #23 and #24. Each fix below was first reproduced with a failing test.

Runtime (Go)

Cancellation no longer aborts an operation that has already started.

• Before: cancellation was checked between every callback and again after each one returned. If the context expired while a successful action ran, SendEvent returned context canceled and the universe stayed in the previous reality, even though the action's side effects had already happened. A retry would repeat them.
• Now: cancellation is honored while waiting for the machine lock and at admission. Once admitted, an operation runs to completion.
• Callbacks still receive the context. One that returns an error, including ctx.Err(), fails like any other callback error.
• An invoke that would start with an already-cancelled context is skipped instead of failing the transition.

MaxConcurrentInvokes now rejects a step before it runs anything.

• Before: invokes were admitted one by one, after the transition actions had run. A rejected invoke aborted the transition while the invokes admitted before it kept running.
• Now: before its first callback, each transition step checks that the pool has room for every invoke it can launch: transition, exit and target-entry invokes, including machine and universe constants. Entries check their entry invokes the same way.
• If the invokes do not fit, the step is rejected with ResourceLimitError before any action runs or invoke starts.

docs/runtime.md and CHANGELOG.md describe the new contract.

Studio

A malformed value or defaultValue no longer crashes the editor.

• The new tryBuildEditorStateFromExternalValue checks that the definition is an object with a universes map, and catches deserialization errors.
• On the first render, the editor falls back to an empty machine. On a controlled update, it keeps the current graph. Both cases report the error with console.error.
• Machines that are valid JSON but have semantic errors still load, so they can be fixed in the editor.

The package styles.css no longer contains @tailwind directives.

• Per the README, the host app's own Tailwind setup generates the utilities.
• Under Tailwind 4, the directives made the package stylesheet emit utilities outside any cascade layer (.transform, .filter).
• I checked the change against the Vite dev server in Chromium.

Gesture checkpoints are cloned once instead of twice.

• A snapshot from createHistorySnapshot is already a detached copy. When it is committed with commit-snapshot, the history now keeps that copy instead of cloning it again.
• Other payloads are still cloned. Callers must not mutate a snapshot after committing it; this is documented in the JSDoc.

Double-clone benchmark (Node 22, editor-core dist, median of 15 runs)

Before is the reducer at fb3d56d and after is this branch; nothing else differs. Each gesture captures a base snapshot, applies 30 silent node moves, then commits once.

Realities	Commit only, 10 gestures (before → after)	Full gesture, 10 gestures (before → after)
100	4.1 ms → 0.5 ms	7.5 ms → 4.0 ms
1,000	35.0 ms → 4.5 ms	68.3 ms → 36.8 ms
5,000	186 ms → 28 ms	362 ms → 206 ms

The commit step is about 7–8× faster, and a full drag gesture about 1.8× faster. The cost that remains is the single clone taken when the gesture starts. A second run gave the same numbers.

Validation

All steps of the Quality workflow were run locally on this branch, and all passed:

• go build ./..., go vet ./..., go test -race -cover ./... (11 packages), make test-fuzz-smoke
• pnpm install --frozen-lockfile, lint, typecheck, test (294 tests), Stryker dry-run, build, audit

The four new Go tests fail against fb3d56d and pass here. The ten new external-value tests failed with TypeError before the fix.