[SandroMaglione]

Problem

Found while reviewing Collision Chamber, a pinball game on Ritmo 0.10. Its adapter runs structuredClone(...) as Ball on every read and write at 240 Hz, because nothing else stops this:

const first = { x: 1 }
// a system: commands.spawn(Game.Command.spawn([Body, first]))
const cp = runtime.checkpoint(simulate)
runtime.tick(simulate)   // a system: set({ x: 2 })
first.x = 777            // host mutates the object it spawned with
runtime.rewind(cp)       // ok: true
// world reads x: 777, not 1

The world stored values by reference, so a later mutation changed the world with no write. changed readers missed it, and rewind brought back the mutated value, breaking the "exactly as it was" promise in Checkpoint.

Change

• A runtime made with debug: { hostClock } deep-freezes every value it is handed, as it is queued or written:

  • Game.Command.spawn / entry / commands.insert
  • a cell's set / update / setRaw, including entity states
  • resource writes, emit, runtime.send, next-state set, machine update
  • the resources and machines given to Game.Runtime.make

  A later mutation throws a TypeError at the line that does it. The walk stops at frozen objects, so { ...value, x } freezes one object. Typed arrays are skipped; Map/Set contents stay writable.

• Release runtimes are unchanged. They freeze nothing, and the write hot path is untouched: debug runtimes get separate cell classes, chosen once when the cell is made.

• Docs:

  • WriteCell.set/update, Command.entry/spawn, ReadonlyValue, and Debug say the world owns what it is given.
  • Runtime.tick points to FixedLoop.start; the game hand-wrote a fixed-step accumulator.
  • The ritmo skill now treats a variable a system closes over like a module variable, which a rewind can't restore.

Breaking for debug runtimes only. The changeset carries the fix: build a new value instead of mutating the one handed over.

Checks

• pnpm run check: passes (1286 tests). test/Runtime.ownership.test.ts covers each entry point, the rewind case, and that release runtimes freeze nothing; 5 of its 8 tests fail without the fix.
• pnpm --dir website build: passes.
• pnpm bench:check fails locally on render/sync-1k-changed-of-10k (critical). main fails it the same way on this machine (+25% over the committed baseline), and main against its own earlier run moved +38–57%, so it is machine noise. Allocations are unchanged.

🤖 Generated with Claude Code