soak: add Gate 8b MAC-churn 24h harness

[lance0]

Summary

Sibling to tests/soak/run-gate8b-soak.sh. Reuses the existing 2-PE shared-ESI topology and layers sustained bridge-FDB churn on top of the DF-flip loop so the soak exercises:

• kernel-learn / local-MAC observation → Type 2 origination
• ADR-0059 receive-side aliasing-ECMP (FDB nexthop groups)
• RFC 7432 §15.1 MAC mobility sequencing
• Gate 8b BUM-suppression while FDB programming is in flight
• ADR-0059 drift-recovery counters under realistic timing

The alpha-checklist exit condition for flipping apply_bum_enforcement and apply_aliasing_ecmp production defaults.

Design choices (per discussion)

• Sibling harness, not a replacement. Base Gate 8b BUM-state soak still useful for memory-only validation.
• Direct kernel FDB mutation (docker exec <pe> bridge fdb add/del <mac> dev ce100a master static). Exercises the local-MAC observation pipeline; deliberately not the gRPC inject path.
• Bounded rotating pool (default 512 MACs, ~256/PE). Per-batch ops every CHURN_INTERVAL_SEC (default 5s), weighted across add / delete / PE-to-PE mobility. Pool occupancy bracketed around target so the soak doesn't drift to empty or saturated.
• DF flips run concurrently. PE2 stop/start clears PE2's pool state file to match kernel reality after restart.
• Sampling additions: kernel-side FDB total + extern_learn counts, ip nexthop count (ADR-0059 NHGs), evpn_local_originations_total, evpn_local_origination_errors_total, evpn_local_observations_dropped_total, evpn_duplicate_mac_moves_total, all four ADR-0059 drift counters, and harness-side add / del / move totals.

Smoke-before-soak path

# 6-min aggressive stress (~25 ops/sec) — catches obvious leaks / hangs.
SOAK_HOURS=0.1 CHURN_INTERVAL_SEC=2 CHURN_BATCH_SIZE=50 \
    bash tests/soak/run-gate8b-mac-churn-soak.sh

# 1h soak — catches slower drift.
SOAK_HOURS=1 bash tests/soak/run-gate8b-mac-churn-soak.sh

# 24h.
bash tests/soak/run-gate8b-mac-churn-soak.sh

Test plan

• bash -n tests/soak/run-gate8b-mac-churn-soak.sh — syntax clean
• 6-minute aggressive stress smoke
• 1h soak
• 24h soak

Notes for review

• No dedicated analyzer yet. tests/soak/analyze-gate8b-soak.py still covers memory / DF gates that apply; MAC-churn-specific gates (e.g. evpn_local_origination_errors_total == 0, drift counters bounded, extern_learn count stable on receiver) surface from manual CSV inspection. A dedicated analyzer can land alongside the first 24h run's postmortem.
• One bug spotted-and-fixed during authoring: the churn loop runs in a background subshell, so the add/del/move totals must live in counter files (not bash vars) for the CSV writer to see them.