PR Summary

Medium Risk
Changes core workload randomness and replay contract (new frozen stream ids); zipfian uses float math and mutex caching where incorrect n stability could hurt throughput, but scope is load-test config rather than auth or payments.

Overview
Replaces stub SampleIndex implementations with real uniform and YCSB zipfian(θ) keyspace index samplers, wired for deterministic replay via Distribution.SetStream and per-scenario RNG sub-streams.

Uniform draws in [0, n) with equal probability (seeded Uint64N or global RNG). Zipfian precomputes zeta(n, θ) once per stable n, caches O(1) draw constants behind a mutex, and skews mass toward low indices as θ rises; n == 0 now errors instead of returning 0 silently. JSON unmarshaling no longer embeds stream state—streams bind at generator startup like gas pickers.

The generator adds bindDistributionStreams for KeyDistribution / SizeDistribution using new append-only frozen stream ids dist:%d:key and dist:%d:size. Package config/doc.go documents wire format, math, and reproducibility contracts. Tests cover determinism, statistical sanity, zipfian performance after warmup, cache invalidation on n change, and numerical edge cases.

^{Reviewed by Cursor Bugbot for commit 3a673df. Bugbot is set up for automated code reviews on this repo. Configure here.}

1. zeta recompute on n change — documented, but worth a runtime warning (or at least a test assertion) if n changes across draws for the same sampler, since the caller contract says it shouldn't.
   Right now a silently-changing n just silently tanks performance with no signal.
2. SampleIndex on UniformDistribution takes a pointer receiver — changed from value receiver in the stub. That's correct (needs to read u.stream), but UniformDistribution doesn't have the copy-unsafe
   warning that ZipfianDistribution does. It's not mutex-bearing so it's technically safe to copy (the stream pointer would alias, not be lost), but adding a note for consistency wouldn't hurt.
3. TestSampleIndexSeededDiffersFromUnseeded — this test uses the global rand RNG for the "unseeded" path. On a fixed seed run the global RNG could theoretically match — though the comment
   acknowledges the probability is ~0. Acceptable in practice.
4. TestZipfianInitCostBounded — the 2-second budget for n=1e6 is very generous. On slow CI machines this is fine, but it means the test won't catch a real performance regression until it's ~100×
   slower than intended. A tighter bound (say 500ms) with a comment justifying it would be more useful as a sentinel.
5. Stream IDs added to the frozen contract — correctly documented in both rng.go and streams.go. The append-only annotation is clear. No concerns.

Thanks @amir-deris — addressed in 9226a2b:

1. zeta recompute on n change — added a test assertion (TestZipfianRecomputesOnNChange) that pins the recompute-on-n-change behavior, rather than a runtime warning. Rationale: config/ is a pure wire-types/library package with zero logging (all log.Printf lives in the service layers — sender/funder/generator/main/stats); adding logging to a sampler hot-path would be the smell you flagged and the only logging in the package. The test locks the contract so a refactor that drops the n-change check is caught. If you'd rather have a runtime signal (e.g. a metrics counter on recompute, since the caller is service-layer), say the word and I'll add it there.
2. UniformDistribution copy note — added for parity: // copy-safe: holds no mutex; the *rng.Stream pointer aliases on copy.
3. TestZipfianInitCostBounded — tightened 2s → 500ms (measured ~30ms init+1000 draws; ~16× headroom, non-flaky under -race, trips well before an O(n)-per-draw regression), with a justifying comment.
4. Points 3 & 5 (unseeded-RNG-match probability ~0; frozen stream-ids) — left as-is per your "acceptable / no concerns."

All green: make lint 0 issues, go test -race -count=5 ./config/... (15 runs) clean. Re-requesting your review since the push dismissed the approval.