fix(object): remove wide-key cutoff and quadratic assign scans

benchmarks/wide_object_membership/README.md

@@ -0,0 +1,129 @@
+# Wide-object membership and `Object.assign` (#10059)
+
+Current main at `603b074ace01464bc66fc07cc8d532f26ccf5a0f` reproduces both reported
+problems. The three TypeScript workloads in this directory are byte-identical
+to the issue attachments. The baseline and fixed sweeps use the same source
+bytes, Node v26.5.1, sequential size order, and matched Perry compiler/runtime
+archives.
+
+## Change
+
+`own_key_present` now delegates to the shared keys-array lookup. A complete
+shape index can prove both presence and absence, so a missing destination key
+in `Object.assign` does not fall through to a scan of every preceding key.
+When the index is missing, shortened, or otherwise incomplete, lookup retains
+the dense-slot fallback. The keys-array length and slot indexes are already
+`u32`, so removing the unrelated 65,536 guard does not widen their
+representation.
+
+`own_key_present_via_index` also consumes the index's explicit `Found`,
+`Absent`, and `Unindexed` verdicts. This preserves the exact fallback for an
+untrusted miss rather than silently treating an index that declined to answer
+as authoritative.
+
+## Method
+
+- Apple M1 Max, 10 logical cores, macOS 26.5, arm64; Node v26.5.1; Perry
+  0.5.1532.
+- Compiler and both matching static archives built together with
+  `CARGO_PROFILE_RELEASE_CODEGEN_UNITS=16 cargo build --release --locked -j 4
+  -p perry -p perry-runtime-static -p perry-stdlib-static`.
+- Each source compiled with `--no-auto-optimize --no-cache` and an explicit
+  matching `PERRY_RUNTIME_DIR`.
+- Sizes run in their issue order. Each process keeps the original warmup (at
+  least 200 ms and five runs), seven samples of at least 20 ms, and a 60-second
+  timeout.
+- The host was busy and load varied materially. Exact constant-factor ratios
+  are descriptive; the checksum repair and slope classification are the
+  useful results.
+
+Baseline load average started/ended at `[11.49, 25.14, 30.11]` / `[24.25,
+25.97, 30.06]`; fixed at `[23.43, 25.68, 29.88]` / `[43.52, 32.26, 31.92]`.
+
+## Membership correctness
+
+Times are median milliseconds. A `no` checksum match is the reported cutoff,
+so that timing is not evidence of correct performance.
+
+| workload | n | Node before | Perry before | checksum | Node after | Perry after | checksum |
+|---|---:|---:|---:|:---:|---:|---:|:---:|
+| `in` | 100 | 0.0038 | 0.0411 | yes | 0.0134 | 0.0680 | yes |
+| `in` | 1,000 | 0.0462 | 0.3969 | yes | 0.0937 | 0.5921 | yes |
+| `in` | 10,000 | 0.6241 | 4.1646 | yes | 1.6149 | 8.1271 | yes |
+| `in` | 100,000 | 8.2211 | 59.4115 | **no** | 22.6291 | 127.2595 | yes |
+| `in` | 1,000,000 | 145.1689 | 653.3018 | **no** | 254.7871 | 2,756.4532 | yes |
+| `hasOwnProperty` | 100 | 0.0069 | 0.0890 | yes | 0.0143 | 0.1259 | yes |
+| `hasOwnProperty` | 1,000 | 0.0780 | 2.5566 | yes | 0.2123 | 1.5102 | yes |
+| `hasOwnProperty` | 10,000 | 0.9632 | 207.5987 | yes | 4.2234 | 15.4140 | yes |
+| `hasOwnProperty` | 100,000 | 14.2578 | 47.8556 | **no** | 39.2262 | 184.3443 | yes |
+| `hasOwnProperty` | 1,000,000 | 442.9680 | 876.0386 | **no** | 546.1051 | 3,357.0082 | yes |
+
+After the fix, all membership checksums match Node through one million keys.
+The fitted Perry-minus-Node exponent delta is `0.061` for `in` and `-0.049`
+for `hasOwnProperty`, consistent with linear work in the number of probes.
+
+The focused boundary fixture separately confirms Node-equivalent results on a
+65,537-key object for present values, present `undefined`, missing and inherited
+keys, short and Unicode names, deletion/reinsertion, and direct reads. On the
+base build, every own membership result becomes false while direct reads remain
+correct.
+
+## `Object.assign`
+
+| n | Node before | Perry before | ratio | Node after | Perry after | ratio |
+|---:|---:|---:|---:|---:|---:|---:|
+| 100 | 0.0298 | 0.1454 | 4.89× | 0.1692 | 0.1790 | 1.06× |
+| 1,000 | 0.3838 | 6.4520 | 16.81× | 0.8589 | 2.9133 | 3.39× |
+| 10,000 | 9.1615 | 528.0727 | 57.64× | 16.3402 | 30.1276 | 1.84× |
+
+The Perry-minus-Node fitted exponent delta changes from **0.536 to 0.121**,
+below the issue's 0.25 target. All copied-value checksums match Node.
+
+The compiled boundary fixture also crosses the former cutoff before exercising
+strict-set behavior. It matches Node byte-for-byte for non-writable data
+properties, getter-only and setter accessors, source getter/setter order,
+symbols, and rejection of new keys on a non-extensible target. The base build
+fails the non-writable-property case because the false membership answer makes
+the precheck treat it as a new property.
+
+## Artifacts and reproduction
+
+The pristine base artifact hashes are:
+
+```text
+perry                  63cb83cf87044443184687ac5eee3376a21dd46c7a1cbf3074789926f99d7b36
+libperry_runtime.a     3c0467d1d7ebcd8bd53e98c43e027ff7e93b306024f68b9ba453f383fb8a6fb4
+libperry_stdlib.a      a13fd68f3b01f814c7fcd031f57b34e978bc00612317f3c20bef68928f7067af
+```
+
+The fixed artifact hashes are recorded in `fixed.json`; its `artifact_source`
+contains the exact runtime diff SHA-256 used for the build. `baseline.json` and
+`fixed.json` retain all medians, run counts, checksums, source hashes, ratios,
+slopes, host metadata, and load averages.
+
+Run a sweep with:
+
+```sh
+python3 benchmarks/wide_object_membership/measure.py fixed \
+  --compiler target/release/perry \
+  --runtime-dir target/release \
+  --artifact-source "$(git rev-parse HEAD)" \
+  --require-checksum-match \
+  --output target/wide-object-membership/fixed
+```
+
+## Validation
+
+- Runtime unit suite: 3,525 passed, 4 ignored, zero failures, single-threaded.
+  This includes the new exact-boundary test and existing object membership,
+  descriptor, enumeration, tombstone, shape-index, and moving-GC coverage.
+- The new 65,537-key compiled fixture matches Node byte-for-byte. A broader
+  object-named parity sweep completed 19/19 fixtures with no mismatches,
+  compilation failures, or crashes before an unrelated ext-wrapper fixture
+  requested its own fresh feature build; the sweep was stopped there to keep
+  the already-validated compiler and archives coherent. The focused fixture
+  and all runtime tests completed independently.
+- Test-registration and GC runtime-root-holder audits pass.
+- `pre-tag-check.sh --quick` passes every check except the pre-existing public
+  benchmark evidence freshness gate. Its registered public source/harness
+  inputs are unchanged by this branch, as are all version and release metadata.