[tobias-tengler]

Stacked on #10653, which adds request-scoped entity reuse to Apollo _entities lookups. This PR targets that branch and merges after it.

A lookup whose entities are all satisfied by the request-scoped entity reuse from #10606 sends no request, but it still collects its targets and producers, writes the variable set JSON for both, and fills every reused field by response name.

The duplicate-target leaf copy from #10612 has the same cost: it resolves every completed leaf by response name for each duplicate target.

This PR matches reuse targets to producers on the raw bytes of their key values and fills the reused fields by row offset, so a fully reused lookup writes no JSON, no variable sets and no representations. The fast path covers composite lookups and Apollo _entities lookups (single and batched) alike; a fully reused Apollo lookup completes without writing a representation and its dependents run. The row offsets are resolved once per plan. If any target has no producer, or the keys are not simple leaf values, nothing is written and the existing path runs unchanged, which for Apollo is the partial reuse of #10653. The duplicate-target leaf copy now assigns leaves by row offset as well. Plans, request counts and results are unchanged.

The numbers below were measured before the fast path was extended to Apollo lookups. They stand for the composite change; Apollo Federation was unchanged at that point because its plans had no entity reuse yet, so only the leaf copy applied. The Apollo extension has not been measured separately.

LookupFanOutBenchmark (added here) runs the whole benchmark query through the gateway executor against canned subgraph responses. Mean of two passes (order A B1 B2, then B2 B1 A), main = A, A1 = reuse fast path, A1+A2 = this PR.

Family	Version	Mean	Ratio	Allocated
Composite	main	301.0 us	1.00	115.55 KB
Composite	A1	289.9 us	0.96	113.85 KB
Composite	A1+A2	275.5 us	0.92	113.85 KB
Apollo Federation	main	319.7 us	1.00	112.25 KB
Apollo Federation	A1	316.2 us	0.99	112.26 KB
Apollo Federation	A1+A2	306.2 us	0.96	112.32 KB

k6 constant, .NET 10, 30 s warm-up and 3 x 30 s, A B B A. Screen holds (composite hold 2, fed hold 1):

Family	Version	RPS	CPU us/req	Alloc KB/req
Composite	main	4870	437	35.7
Composite	PR	4874	420	33.85
Apollo Federation	main	4808	446	42.05
Apollo Federation	PR	4773	446	41.9

Confirmation runs (one hold per cell, run CV in parentheses). Cells marked noisy have a run CV above 5% on at least one side and are not conclusive.

Cell	Version	RPS	CPU us/req	Alloc KB/req
Composite constant, A B (noisy)	main	4557 (6.0%)	442	35.7
Composite constant, A B (noisy)	PR	4434 (6.2%)	487	34.0
Composite constant, B A	main	4845 (3.2%)	439	35.7
Composite constant, B A	PR	4767 (2.4%)	414	33.9
Composite constant-latency, B A	main	1967 (0.6%)	504	35.9
Composite constant-latency, B A	PR	1974 (2.4%)	474	34.2
Composite burst, A B (noisy, outlier)	main	5183 (2.0%)	474	36.5
Composite burst, A B (noisy, outlier)	PR	4285 (17.5%)	534	34.6
Apollo Federation constant, B A (noisy)	main	4684 (6.9%)	460	41.9
Apollo Federation constant, B A (noisy)	PR	4364 (4.5%)	469	41.9

The stable signal is allocation on composite: -4.7% to -5.3% KB/req in every pair. CPU/req is -4% to -6% in the clean cells (screen hold, constant B A, constant-latency); the noisy cells (constant A B, burst, Apollo Federation constant) show no pattern that can be attributed to the change and were not repeated cleanly. RPS does not move because the composite scenario is not gateway-bound at this load. Apollo Federation was unchanged in these runs: its plans had no entity reuse before #10653, so only the leaf copy applied and it is below the noise. Harness noise is about 3% RPS and 1.5% CPU/req for composite, 2% and 1% for Apollo Federation.

[github-actions]

Patch coverage

85.6% of changed lines covered (370/432)

File	Covered	Changed	Patch %
…/Execution/Results/EntityReuseCandidateMap.cs	47	59	79.7% 🔴
…/Execution/Results/FetchResultStore.Pooling.cs	16	20	80.0% 🟡
…/src/Fusion.Execution/Execution/Results/FetchResultStore.cs	238	278	85.6% 🟡
…/src/Fusion.Execution/Text/Json/CompositeResultDocument.cs	45	51	88.2% 🟡
…/Execution/Results/EntityReuseCandidateList.cs	3	3	100.0% 🟢
…/src/Fusion.Execution/Execution/Results/EntityReuseLayout.cs	10	10	100.0% 🟢
…/Execution/Results/FetchResultStore.RepresentationValue.cs	11	11	100.0% 🟢

Uncovered changed lines (JSON)

{
  "sha": "f145f066a80131e6c69bc714c1d98f8ce0e79c8c",
  "files": [
    { "path": "src/HotChocolate/Fusion/src/Fusion.Execution/Execution/Results/EntityReuseCandidateMap.cs", "ranges": [[326, 327], [333, 334], [344, 345], [349, 350], [370, 371], [386, 387]] },
    { "path": "src/HotChocolate/Fusion/src/Fusion.Execution/Execution/Results/FetchResultStore.Pooling.cs", "ranges": [[226, 229]] },
    { "path": "src/HotChocolate/Fusion/src/Fusion.Execution/Execution/Results/FetchResultStore.cs", "ranges": [[897, 898], [909, 910], [983, 985], [991, 993], [1309, 1310], [1314, 1315], [1324, 1324], [1330, 1331], [1344, 1346], [1365, 1366], [1372, 1373], [1377, 1380], [1404, 1405], [1415, 1416], [1462, 1462], [1464, 1465], [1508, 1509], [1513, 1515]] },
    { "path": "src/HotChocolate/Fusion/src/Fusion.Execution/Text/Json/CompositeResultDocument.cs", "ranges": [[736, 737], [759, 762]] }
  ]
}

---

Project coverage: 59.3% (314853/531207 lines)