[g4lb]

scope.drafts_ now holds the draft states, not the draft proxies. Before this change, revokeDraft read draft[DRAFT_STATE], and each read ran the proxy get trap. Now revokeDraft gets the state directly. The output does not change.

Built dist/cjs production code, ns/op:

Benchmark	Before	After	Change
produce on a small object with no writes	381	345	-9.5%
Update 2 nested values with autoFreeze: false	1,983	1,796	-9.5%
Update 1,000 objects in an array	633,221	584,852	-7.6%
Read 100 array items with no writes	125,348	117,060	-6.6%
Update 25 records of mixed shapes	33,163	32,239	-2.8%

I measured no slowdown on 29 benchmarks. These include Map and Set drafts, patches, createDraft and finishDraft, and array methods. Some small benchmarks change by 10-20% between rounds. On those benchmarks, I measured no significant change. I used 12 interleaved rounds and a Mann-Whitney test (Apple M2 Max, Node 24.15.0). I did not measure the speed on other engines.

Checks:

• yarn test passes.
• Old and new outputs matched on 1,000,000 random recipes in Node. The comparison includes results, structural sharing, frozen state, patches, and errors.
• Old and new built outputs matched on 50,000 random recipes each in Bun and Deno.

[markerikson]

Nice! What happens if you run the existing benchmarks suite (possibly with an update to include existing Immer 11.x as a comparison point)?