perf(feature-flagging): avoid duplicate exposure context capture

[leoromanovsky]

Motivation

The evaluator copied the complete context before the writer removed duplicate exposures. Repeated exposure cost increased with customer context size. A 1,000-field duplicate allocated 162,000 bytes and took 29.65 microseconds.

Changes and Decisions

• Check the four-field exposure identity before context capture.
• Keep the serializer LRU cache as the final exact deduplication guard.
• Record admission only after the queue accepts an event.
• Bound the admission cache to the queue capacity.
• Replace snapshot-then-flatten with one bounded context copy.
• Preserve the targeting key and normal exposure payload semantics.
• Add no worker or scheduler thread.

Validation

JDK 11 JMH, one thread:

Exposure path	Before	After
Duplicate, empty	81.07 ns, 1,264 B/op	7.31 ns, 24 B/op
Duplicate, 100 flat fields	2.77 us, 18,032 B/op	7.38 ns, 24 B/op
Duplicate, 100 nested fields	8.90 us, 42,952 B/op	7.72 ns, 24 B/op
Duplicate, 1,000 flat fields	29.65 us, 162,000 B/op	7.77 ns, 24 B/op
First, 1,000 flat fields	30.62 us, 162,000 B/op	3.91 us, 13,128 B/op

[datadog-datadog-prod-us1]

🎯 Code Coverage (details)
• Patch Coverage: 97.70%
• Overall Coverage: 78.34% (+19.99%)

_{This comment will be updated automatically if new data arrives.
🔗 Commit SHA: 577b37b | Docs | Datadog PR Page | Give us feedback!}