Answer a wrapped dictionary's existence questions from ContainsKey

[lahma]

Draft-quality until its gate lands: the benchmark tables (new InteropDictionaryProbeBenchmark + controls + full SunSpider/Dromaeo) are being measured now and will be posted here before this should merge.

ObjectWrapper had no ProbeOwnProperty override, so every existence question — in, hasOwnProperty, and the per-key filter behind Object.keys/for-in/spread/JSON.stringify — fell to the base implementation, whose answer is the read: TryGetValue, FromObject on the value (for an object value, a whole nested ObjectWrapper), a PropertyDescriptor around it — then both discarded. Dictionary members are the one thing the wrapper never caches, so this repeated on every call: Object.keys over a wrapped JSON document paid it once per key. #2839 compiled ContainsKey, but only HasProperty's non-string-keyed arm could reach it — the one shape whose keys never enumerate.

The override answers Enumerable on a ContainsKey hit and otherwise falls through to base — it never decides a key is missing, because a wrapped dictionary still resolves CLR members for names the dictionary does not carry. The only possible disagreement is a target whose own ContainsKey and TryGetValue disagree, which the compiled lanes already trust. Precedence mirrors GetOwnProperty's own order. TypeDescriptor.CanTestDictionaryKey exists because ContainsDictionaryKey returns false both for "absent" and for "there was no ContainsKey to ask" — and reading the second as the first is exactly the silent key-dropping the probe contract warns against.

Proof it engages: a counting IDictionary shows exactly one ContainsKey and zero TryGetValue for 'a' in doc. The in-box agreement checker covers plain keys, object-valued keys, a key shadowing a CLR member, a CLR member name, an absent name, an index-shaped name, a symbol, and a script-redefined non-enumerable key; the embedder-side half lives in Jint.Tests.PublicInterface/HostDictionaryEnumerationTests.cs with the host dictionary as oracle. Green under JINT_HOST_CONTRACT_VERIFICATION=1 on both projects — which verifies every consumed probe against GetOwnProperty — and AGENTS.md's dictionary gotcha records the new lane.

The new benchmark's own InMiss row is the honest control: a miss can only get slower (one extra compiled ContainsKey on top of the resolver lookup it already paid), and by how much is a number the gate reports rather than hides. ObjectValues and ReadMember are value-needing controls no existence answer can serve.

Full solution green on both TFMs; test262 exactly 99,738 / 0 / 157.

🤖 Generated with Claude Code

[lahma]

Benchmark gate

Targeted — InteropDictionaryProbeBenchmark (new in this PR) vs the merge-base 20bbf8725:

Row	base	cand	time Δ	alloc
InHit	74.2 µs	49.8 µs	−32.8%	31.86 KB → 624 B (−98.1%)
HasOwnProperty	191.6 µs	153.6 µs	−19.8%	31.86 KB → 624 B (−98.1%)
ObjectKeys	824.1 µs	520.5 µs	−36.8%	−31.8%
ForIn	1.656 ms	1.320 ms	−20.3%	−37.2%
ObjectValues	1.228 ms	1.020 ms	−16.9%	−29.5%
InMiss (cost-of-asking control)	77.5 µs	83.1 µs	+7.2%	flat
ReadMember (control)	111.8 µs	109.3 µs	−2.2%	flat

The allocation column is the byte-deterministic channel and confirms every designed row: the discarded per-probe descriptor and nested wrapper are gone. ObjectValues improves despite being listed as a value-needing control, mechanically: its enumerability filter no longer materializes a descriptor per key before the values are read. InMiss +7.2% is the predicted, disclosed cost — one extra compiled ContainsKey ahead of the resolver lookup a miss already paid, ≈ 5.6 ns per miss.

Neighbours: InteropNestedDictionaryBenchmark and ImmutableCrossingBenchmark (whose read paths this diff does not enter) showed one +11.0% flag in pair 1; the order-reversed second pair exonerated it — every row within ±2.4%, allocation byte-flat in every run.

Wide — SunSpider (26) + Dromaeo (24), single pair: allocation flat on all 50 rows. No SunSpider or Dromaeo script wraps a CLR dictionary, so no row executes the new code; the time swings (both directions, up to ±12%) appear on exactly those unreachable rows and contradict the movements the unrelated #2968 candidate showed on the same rows against the same baseline — single-draw baseline noise, reported as such.

Machine idle, serial, default BDN job, both sides from their own worktree CWD; the new benchmark class was ported byte-identical to the baseline side so both ran the same harness.