[ejc3]

Fixes #1619.

WeakMap and WeakSet found a key by walking every entry of the collection (scr_weak_find over map->entries in packages/runtime/src/scr_json.c), and scr_weak_remove_impl walked the list again to unlink an entry, so filling a collection with n keys took O(n²) time. 200,000 object keys took 64.6 s on Linux arm64 where Node takes 0.12 s.

• Each collection keeps its entries in a hashed index (buckets, a power-of-two length) that doubles when its entries outnumber its buckets. get, set, has and delete look up and unlink through it in expected constant time, and tracing, release and collection walk its buckets.
• The thread-wide index of all entries, which key disposal searches, starts from the same static 1,024 buckets, moves to a heap table that doubles as entries grow, and returns to the static table when the last entry goes, so a thread that ends with no weak entries frees it.
• Key disposal returns at once when no weak entry exists.
• Both indexes hash the pointer with a multiplicative mix (scr_weak_hash), which spreads keys over any table size; the old (bits >> 4) ^ (bits >> 14) was fitted to a fixed 1,024-bucket table.

Tests:

• Corpus 1619 covers both indexes growing past their first sizes, overwrites, deletes and re-adds, keys shared by two maps and a set, a map dropped while its keys live, keys dropped while maps live, and fresh keys (which may reuse a dropped key's memory) not being found.
• tests/harness/weak-collection-scaling.test.ts runs the same work with 25,000 and 200,000 keys and bounds the ratio of the two times at 32. With this change it measures 11x to 14x, in both lanes; on main the 200,000-key run is stopped at its 60 s timeout. It follows closure-environment-scaling.test.ts.

Testing

On main at 1ee5c84. Environment: Linux arm64 (glibc), Node 24.22.0, pnpm 11.1.3, LLVM 22.1.8, Zig 0.16.0, with the llvm-linux-arm64-gnu and runtime-linux-arm64-gnu native artifacts built locally from this tree. I have no Vercel Sandbox credentials, so these are the local fallback lanes from CONTRIBUTING.md. The host was shared with other test runs.

• The issue's program (200,000 keys) runs in 0.06 s with this branch, against 0.11 s for Node and 64.6 s for 0.4.0. Its output matches Node's.
• Red on main: weak-collection-scaling.test.ts fails at its 60 s timeout on the 200,000-key run. With this branch it passes, plain and with SCRIPTC_SAN=1. Corpus 1619 passes in differential.test.ts and llvm-differential.test.ts, plain and sanitized.
• SCRIPTC_TEST_WORKERS=4 pnpm test: 262 failed, 12620 passed, 126 skipped; main gives 263 failed, 12616 passed on this host, one of them this branch's scaling test, which I had copied there to see it fail. No test fails on this branch only, and library-contract.test.ts's identity-hash case failed on main only. The shared failures: those that target wasm32-wasi (wasi-differential, three, static-spatial, wasm-library, library-dispatch), since no wasm32-wasi runtime pack is built here; 40 in native-toolchain.test.ts, most with ENOENT on scratch cache directories; the 600 s timeout of coverage.test.ts's all-static check; the size bound in runtime-tree-shaking.test.ts; and one ASan case in library-mode.test.ts.
• SCRIPTC_SAN=1: a full local lane with four workers would have taken more than a day here, so I ran it in two parts on this head and on main. The corpus files (differential.test.ts and llvm-differential.test.ts, in 12 SCRIPTC_TEST_SHARD shards): 7016 passed, 0 failed (main: 7014 passed, 0 failed). Everything else (--exclude those two, 12 workers): 266 failed, 5571 passed, 155 skipped; main gives 266 failed, 5570 passed, and the failing tests are the same: the ones above, plus three ASan cases in library-multi.test.ts and one in self-hosting-native-driver.test.ts.
• pnpm lint, pnpm format:check and pnpm test:ts7 --baselines-only pass.
• Not run: the Sandbox gate, macOS, Windows, and other targets' runtime packs (only linux-arm64-gnu was rebuilt).

[vercel]

@ejc3 is attempting to deploy a commit to the Vercel Labs Team on Vercel.

A member of the Team first needs to authorize it.

[ejc3]

Main has moved since the Testing section's 1ee5c84. The two commits apply cleanly to 7c235ce, and there the scaling test, native-weak-collections.test.ts and the weak corpus programs in differential.test.ts and llvm-differential.test.ts pass (plain lane, 33 tests). I have not rerun the full lanes on 7c235ce.