tune: fake-exporter cardinality down (5K aggregate target, was 25K+)

[zzylol]

Summary

The post-PR-#338 five-sketch workload was emitting ~2.46 MB/s SDK output, blowing up agent -> gateway bandwidth and amplifying every other bug. The HLL inner-label fan-out (user_id x outer label set) was the dominant contributor. This PR lowers two binary defaults so the demo lands at the documented ~5K aggregate-series gateway target.

Changes (current -> new)

knob	file	current default	new default	rationale
EXPORTER_FIVE_SKETCH_USER_POOL	deploy/fake-exporter/five_sketch_workload.go	1000 (floor 500)	100 (floor 50)	HLL still has a non-trivial active-user cardinality to estimate; per-min rotate semantic intact.
EXPORTER_CARDINALITY	deploy/fake-exporter/main.go	1000	500	Direct-binary use now matches run_mvp_demo.sh's PER_AGENT_CARDINALITY=500 default; N_PRODUCERS=10 x 500 = 5K aggregate.
EXPORTER_FIVE_SKETCH_ENDPOINTS	deploy/fake-exporter/five_sketch_workload.go	50	50 (unchanged)	Already small; Zipfian heavy hitters dominate top-K anyway.
request_size_bytes	scalar	n/a	n/a	No cardinality knob.

Aggregate gateway cardinality (math)

• 10 producers x PER_AGENT_CARDINALITY=500 outer label sets = 5,000 outer series (target).
• HLL unique_users_per_min: ~5,000 outer x ~100 active user ids per rotation window, with rotation moving the active window through the pool — bounded inflation.
• CountSketch / CountMinSketch endpoints: 5,000 x ~50 = ~250K event keys but Zipfian heavy hitters concentrate on ~5 endpoints, so series count is bounded by 50 endpoint suffixes per outer.
• KLL request_size_bytes: scalar per outer = 5,000 series.

Pre-tune we were at 10K-25K+ series; this brings the gateway closer to the runbook's 5K target line.

Driver / config sanity

• deploy/scripts/run_mvp_demo.sh line 111: PER_AGENT_CARDINALITY=500 already matches.
• deploy/docker-compose/mvp-multi-stage.yml: per-producer env propagates PER_AGENT_CARDINALITY:-500. No override above 500.
• deploy/configs/mvp-workload.yaml: no cardinality pin.
• EXPORTER_FIVE_SKETCH_USER_POOL / _ENDPOINTS are not set externally anywhere — the new binary defaults take effect.

Don't / scope guard

• Metric SHAPES preserved: still request_size_bytes (KLL gauge), unique_users_per_min (HLL counter w/ user_id), top_endpoint_qps / endpoint_request_freq (CountSketch / CMS counter w/ endpoint).
• gorillas3processor untouched (parallel agent owns OOB fix).
• Static placeholder yaml untouched (parallel agent).

Test plan

• go test ./deploy/fake-exporter/... — pass (all 4 metrics still emit; existing test uses explicit USER_POOL=500 so floor change is non-breaking).
• go build ./deploy/fake-exporter/... — pass.
• Local binary smoke — startup log confirms cardinality=500 users=100 endpoints=50 outer_card=500 (was cardinality=1000 users=1000).
• Full demo run (deferred — out of scope per task instructions).

Refs #46.

Generated with Claude Code (1M context).