fix: make scaling_decisions cumulative via Redis sorted set history

[sylvesterdamgaard]

Summary

Resolves #23 — scaling_decisions[] in the cluster summary was per-cycle (rebuilt every ~5s evaluation tick, lost when activity stops). Dashboard cards like "Total Decisions" showed 0 on idle clusters that had been scaling for hours.

• Persist decisions to Redis sorted set — Every non-hold scaling decision is zadd'd to queue-autoscale:cluster:{app-id}:decisions:history with microtime(true) as score
• Dual-bound pruning — Time-based (zremrangebyscore, default 1 hour) + count-based (zremrangebyrank, default 10k entries) prevents unbounded growth
• Atomic writes — Record + prune uses a single Lua eval to minimize Redis round-trips
• Rolling history in summary — buildClusterSummary now reads from recentDecisions() instead of the ephemeral per-cycle array
• Configurable — decision_history_seconds and decision_history_max config keys with env var support

No API break. scaling_decisions[] keeps the same structure, just gains a recorded_at float and historical entries. queue-monitor dashboard cards will populate immediately on upgrade — no consumer-side changes needed.

Test plan

• 10 new unit tests for ClusterStore::recordDecision() / recentDecisions() (record, retrieve, empty state, timestamp, accumulation, time pruning, count pruning, key pattern, score range, malformed JSON, Lua atomicity)
• 1 integration test verifying historical decisions flow through buildClusterSummary
• Full suite: 560 tests, 1667 assertions passing
• PHPStan clean
• Pint clean

🤖 Generated with Claude Code