feat: evaluate big segments

[beekld]

Evaluate big segments

Wires the big segment store wrapper into the server-side evaluator so flags referencing big segments actually evaluate, and annotates the reason with the big segments status. Builds on SDK-2366, which introduced the EvaluationReason::BigSegmentsStatus type.

What's in

• Big segment matching in rules.cpp:
  • looks up membership for the unbounded context kind
  • tests the <key>.g<generation> segment ref
  • short-circuits to non-match on a store error
  • falls through to the segment's rules on a membership miss
• EvaluationStack carries the per-evaluation status, membership cache, and per-key store-error set, so a context key is queried at most once per evaluation.
• The aggregated status is stamped onto the returned EvaluationReason.

Design notes

When one evaluation queries multiple big segments with differing statuses, the least-trustworthy status wins (NOT_CONFIGURED > STORE_ERROR > STALE > HEALTHY). This follows the spec but diverges from Java/Go, which take the last status written.

Testing

Added unit tests covering each evaluation branch against an in-memory store.

---

Note

Medium Risk
Changes core segment-matching and evaluation-reason behavior for flags using big segments; incorrect membership or status aggregation could change targeting outcomes, though behavior is heavily unit-tested.

Overview
Big segment evaluation is wired into the server-side flag evaluator: unbounded segments now resolve membership via BigSegmentStoreWrapper instead of always returning non-match with a TODO.

EvaluationStack is extended with a non-owning store pointer, per-evaluation membership cache (one store lookup per context key), per-key store-error short-circuit, and aggregated BigSegmentsStatus using worst-wins precedence (NOT_CONFIGURED > STORE_ERROR > STALE > HEALTHY). Evaluator accepts an optional big-segment store, seeds the stack from it, and stamps the aggregated status onto the returned EvaluationReason when any big segment was consulted.

In Contains / MatchBigSegment, big segments use the unbounded context kind’s key, check <segmentKey>.g<generation>, treat store errors as non-match without retry, skip regular include/exclude lists, and fall through to segment rules on membership miss. New gtest coverage exercises not-configured, healthy/stale/error paths, precedence, caching, and rule fallthrough.

^{Reviewed by Cursor Bugbot for commit 9b37b9a. Bugbot is set up for automated code reviews on this repo. Configure here.}

[kinyoklion]

nit: Part of me wants this to be more colocated with the evaluation. But I suppose a real constructor change would be noticed.