1 files changed, 133 insertions, 0 deletions
diff --git a/prototypes/workbench/evaluation.test.mjs b/prototypes/workbench/evaluation.test.mjs
new file mode 100644
index 0000000..21a9590
--- /dev/null
+++ b/prototypes/workbench/evaluation.test.mjs
@@ -0,0 +1,133 @@
+import test from 'node:test';
+import assert from 'node:assert/strict';
+import { readFile } from 'node:fs/promises';
+import { createRuntime } from './web/runtime.mjs';
+
+const wasm = await readFile(new URL('./web/core.wasm', import.meta.url));
+const make = (options = {}) => createRuntime(wasm, options);
+const source = `state x = 3
+state spare = 0
+input next = KeyN
+input other = KeyO
+input fail = KeyF
+text value(0, 0, (x + 1) * (x + 2) / 2)
+on next.down { x = x + 0.25 }
+on other.down { spare = spare + 1 }
+on fail.down { x = 900; spare = 1 / 0 }
+`;
+
+test('cache reuse, unrelated changes and discarded records preserve authored values', async () => {
+  const cached = await make({ cacheEntries: 32 }), plain = await make();
+  cached.apply(source); plain.apply(source);
+  assert.deepEqual(cached.snapshot(), plain.snapshot());
+  assert.deepEqual(cached.snapshot(), plain.snapshot());
+  assert(cached.evaluationStats().cacheHits > 0);
+  assert(cached.evaluationStats().instructions < plain.evaluationStats().instructions);
+  for (let i = 0; i < 100; i++) {
+    const name = i % 3 ? 'other' : 'next';
+    for (const runtime of [cached, plain]) { runtime.down(name); runtime.up(name); }
+    if (i % 7 === 0) cached.forgetEvaluation('history');
+    if (i % 11 === 0) cached.forgetEvaluation('values');
+    if (i % 13 === 0) cached.forgetEvaluation('all');
+    assert.deepEqual(cached.snapshot(), plain.snapshot());
+  }
+  // Omit observations long enough to overwrite every change record.
+  for (let i = 0; i < 20; i++) for (const runtime of [cached, plain]) {
+    runtime.down('next'); runtime.up('next');
+  }
+  assert.deepEqual(cached.snapshot(), plain.snapshot());
+});
+
+test('committed changes only: failed events and failed edits preserve reusable output', async () => {
+  const cached = await make({ cacheEntries: 32 }); cached.apply(source);
+  const before = cached.snapshot();
+  cached.down('fail'); cached.up('fail');
+  const failed = cached.snapshot();
+  assert.deepEqual(failed.states, before.states);
+  assert.deepEqual(failed.views, before.views);
+  assert.equal(failed.lastError, 4);
+  assert.equal(cached.evaluationStats().instructions, 0);
+  assert.throws(() => cached.apply('state x = missing'));
+  assert.deepEqual(cached.snapshot(), failed);
+  assert.equal(cached.evaluationStats().instructions, 0);
+});
+
+test('program edits, state reordering and reset cannot reuse old definitions', async () => {
+  const cached = await make({ cacheEntries: 32 }), plain = await make();
+  for (const r of [cached, plain]) { r.apply(source); r.down('next'); r.up('next'); r.snapshot(); }
+  const changed = source.replace('state x = 3\nstate spare = 0', 'state spare = 9\nstate x = 300')
+    .replace('(x + 1) * (x + 2) / 2', '(x + 10) * (x + 20) / 2');
+  for (const r of [cached, plain]) r.apply(changed);
+  assert.deepEqual(cached.snapshot(), plain.snapshot());
+  for (const r of [cached, plain]) r.reset();
+  assert.deepEqual(cached.snapshot(), plain.snapshot());
+  assert.equal(cached.snapshot().states.x, 300);
+});
+
+test('finite float semantics include signed zero, high state bit and arithmetic errors', async () => {
+  const cached = await make({ cacheEntries: 32 }), plain = await make();
+  const program = `${Array.from({ length: 32 }, (_, i) => `state s${i} = 0`).join('\n')}
+input flip = KeyF
+text sign(0, 0, s31)
+text error(0, 0, 1 / s31)
+on flip.down { s31 = -s31 }
+on flip.up { s31 = 2 }
+`;
+  for (const r of [cached, plain]) r.apply(program);
+  assert.deepEqual(cached.snapshot(), plain.snapshot());
+  for (const r of [cached, plain]) r.down('flip');
+  assert.deepEqual(cached.snapshot(), plain.snapshot());
+  assert(Object.is(cached.snapshot().views[0].value, -0));
+  assert.equal(cached.snapshot().displayError.code, 100);
+  for (const r of [cached, plain]) r.up('flip');
+  assert.deepEqual(cached.snapshot(), plain.snapshot());
+  assert.equal(cached.snapshot().displayError, null);
+});
+
+test('one budget covers the whole scene; failures publish no partial scene or state rollback', async () => {
+  const r = await make({ cacheEntries: 32 }); r.apply(source);
+  r.down('next'); r.up('next');
+  const good = r.snapshot();
+  r.configureEvaluation({ workLimit: 5 });
+  const limited = r.snapshot();
+  assert.equal(limited.lastError, 101);
+  assert.deepEqual(limited.views, []);
+  assert.deepEqual(limited.states, good.states);
+  assert.equal(limited.successfulEvents, good.successfulEvents);
+  assert(r.evaluationStats().work <= 5);
+  r.down('fail'); r.up('fail');
+  assert.equal(r.snapshot().lastError, 4);
+  assert.equal(r.snapshot().displayError.code, 101);
+  r.configureEvaluation({ workLimit: 2048 });
+  assert.deepEqual(r.snapshot().views, good.views);
+  assert.equal(r.snapshot().lastError, 4);
+});
+
+test('small caches, no history and every small work limit give exact results or explicit failure', async () => {
+  const plain = await make(); plain.apply(source);
+  for (const cacheEntries of [0, 1, 32]) for (const historyEntries of [0, 1, 8]) {
+    const r = await make({ cacheEntries, historyEntries }); r.apply(source);
+    for (let workLimit = 0; workLimit < 100; workLimit++) {
+      r.configureEvaluation({ workLimit });
+      const actual = r.snapshot();
+      assert(r.evaluationStats().work <= workLimit);
+      if (actual.lastError) {
+        assert.equal(actual.lastError, 101);
+        assert.deepEqual(actual.views, []);
+      } else assert.deepEqual(actual, plain.snapshot());
+    }
+  }
+});
+
+test('invalid settings and forget names preserve policy and state', async () => {
+  const r = await make({ cacheEntries: 32 }); r.apply(source);
+  const before = r.snapshot(), policy = r.evaluationStats();
+  for (const option of [{cacheEntries:33}, {historyEntries:9}, {workLimit:-1},
+    {workLimit:0x100000000}, {workLimit:NaN}, {cacheEntries:0.5}]) {
+    assert.throws(() => r.configureEvaluation(option));
+    assert.deepEqual(r.evaluationStats(), policy);
+  }
+  for (const name of ['unknown', '__proto__', 'constructor']) assert.throws(() => r.forgetEvaluation(name));
+  assert.deepEqual(r.snapshot(), before);
+  assert.equal(r.snapshot().memory.linearBytes, 65536);
+});