diff --git a/packages/cachekit/src/serialization/interop.test.ts b/packages/cachekit/src/serialization/interop.test.ts
index d35b01a..27fde92 100644
--- a/packages/cachekit/src/serialization/interop.test.ts
+++ b/packages/cachekit/src/serialization/interop.test.ts
@@ -138,6 +138,55 @@ describe('interop argument encoding (args profile)', () => {
   });
 });
 
+describe('interop Set encoding budgets (encodeCanonical, shared by both profiles)', () => {
+  it('rejects a Set whose elements are collectively over budget during iteration', () => {
+    // 32 × ~64 KiB elements: each far under the 1 MiB budget, ~2 MiB in
+    // aggregate. Set element sub-encodes buffer into `encoded[]` before the
+    // parent sink sees any bytes, so the running total must be threaded
+    // across the loop — otherwise all 32 elements materialise before the
+    // budget fires. Getter spies count how many elements were actually
+    // encoded: each element encodes to 65,550 bytes, so the throw lands
+    // during element 16 — mid-iteration, not after the loop.
+    const total = 32;
+    let encoded = 0;
+    const elements = Array.from({ length: total }, (_, i) => ({
+      get payload(): string {
+        encoded++;
+        return `${i}:`.padEnd(64 * 1024, 'x');
+      },
+    }));
+    expect(() => encodeInteropValue(new Set(elements))).toThrow(ValueTooLargeError);
+    expect(encoded).toBeLessThan(total);
+  });
+
+  it('accepts a duplicate-heavy Set whose deduped encoding fits the budget', () => {
+    // 20 distinct-identity objects with identical canonical encodings: ~4 MiB
+    // pre-dedupe, ~200 KiB deduped. Dedupe happens on insert, so duplicates
+    // advance neither the byte budget nor the count — this must encode
+    // byte-identically to the singleton, not throw at the pre-dedupe sum.
+    const dup = (): { k: string } => ({ k: 'x'.repeat(200 * 1024) });
+    const many = new Set(Array.from({ length: 20 }, dup));
+    expect(many.size).toBe(20);
+    expect(encodeInteropValue(many)).toEqual(encodeInteropValue(new Set([dup()])));
+  });
+
+  it('rejects a Set with too many distinct elements during iteration', () => {
+    // 10,002 tiny distinct elements: far under the byte budget, over the
+    // 10,000 collection cap. The cap fires on the 10,001st retained element,
+    // not after the full Set has been encoded and buffered.
+    const total = 10_002;
+    let encoded = 0;
+    const elements = Array.from({ length: total }, (_, i) => ({
+      get n(): number {
+        encoded++;
+        return i;
+      },
+    }));
+    expect(() => encodeInteropValue(new Set(elements))).toThrow(ValueTooLargeError);
+    expect(encoded).toBeLessThan(total);
+  });
+});
+
 describe('interop value encoding (value profile)', () => {
   it('maps undefined to nil (no cross-SDK arity contract for values)', () => {
     expect(hex(encodeInteropValue(undefined))).toBe('c0');
diff --git a/packages/cachekit/src/serialization/interop.ts b/packages/cachekit/src/serialization/interop.ts
index 89d6f32..215015d 100644
--- a/packages/cachekit/src/serialization/interop.ts
+++ b/packages/cachekit/src/serialization/interop.ts
@@ -126,7 +126,11 @@ function concatChunks(chunks: Uint8Array[]): Uint8Array {
  */
 interface ChunkSink {
   chunks: Uint8Array[];
-  /** Running encoded byte count (starts above 0 for Set element sub-encodes). */
+  /**
+   * Running budget cursor. Set element sub-encodes seed it with the parent
+   * total plus prior retained elements' bytes, so it may exceed the total
+   * length of `chunks`.
+   */
   bytes: number;
 }
 
@@ -410,19 +414,32 @@ function encodeCanonical(
     // encoded bytes (unsigned lexicographic) and dedupe post-normalization —
     // a total, language-neutral order (spec: "Set ordering is not numeric
     // order").
+    // Elements encode into isolated sub-sinks (the byte-order sort needs each
+    // element's bytes) threaded through a running byte total, and dedupe
+    // happens on insert, so the byte budget and the collection-size cap both
+    // fail DURING iteration — before every element is encoded and buffered —
+    // while counting exactly what the output retains: duplicates advance
+    // neither total. Acceptance matches checking after the loop, except that
+    // a duplicate arriving with less than its own encoded size left in the
+    // byte budget still throws mid-encode (duplicateness is unknowable until
+    // encoded). The parent's own total advances once, on the pushes below.
     const encoded: Uint8Array[] = [];
+    const seen = new Set<string>();
+    let running = sink.bytes;
     for (const element of v) {
-      // Isolated sub-sink (the byte-order sort needs each element's bytes),
-      // inheriting the parent's running total so a huge element fails fast.
-      // Parent bytes are counted once, when the deduped buffers are pushed.
-      const sub: ChunkSink = { chunks: [], bytes: sink.bytes };
+      const sub: ChunkSink = { chunks: [], bytes: running };
       encodeCanonical(element, profile, depth + 1, sub);
-      encoded.push(concatChunks(sub.chunks));
+      const bytes = concatChunks(sub.chunks);
+      const key = bytesToHex(bytes);
+      if (seen.has(key)) continue;
+      seen.add(key);
+      checkCollectionSize(seen.size, 'array');
+      running = sub.bytes;
+      encoded.push(bytes);
     }
     encoded.sort(compareBytes);
-    const deduped = encoded.filter((b, i) => i === 0 || compareBytes(b, encoded[i - 1]!) !== 0);
-    encodeArrayHeader(deduped.length, sink);
-    for (const b of deduped) pushChunk(sink, b);
+    encodeArrayHeader(encoded.length, sink);
+    for (const b of encoded) pushChunk(sink, b);
   } else if (Array.isArray(v)) {
     encodeArrayHeader(v.length, sink);
     for (const item of v) {