fix: keep Stream metric scrapes off durable storage

[smiggleworth]

Summary

• maintain Stream event and watermark metrics in a monotonic in-process projection
• initialize the projection from persisted state at startup and advance it only after successful durable commits
• render raw Prometheus metrics without storage scans or actor mailbox work while preserving refreshed structured admin metrics
• yield Stream maintenance after one bucket so synchronous family actors service client commands between strict storage commits
• pin the latest Midge main revision with bounded-memory compaction and update the auto-memory expectation

Why

Fitz Destroyer sustained pressure showed that repeated /metrics scrapes scanned growing Stream history and competed with data-plane storage and actor work. After removing that scrape path, sustained pressure exposed multi-bucket Stream maintenance monopolizing the same synchronous family actor. This keeps both observability and maintenance within the documented liveness boundary.

Validation

• cargo fmt --all -- --check
• cargo test --workspace --quiet (1,615 unit tests plus all integration and doc tests)
• cargo clippy --locked --workspace --all-targets --all-features -- -D warnings -D clippy::pedantic
• Fitz Destroyer eight-minute soak, four clients, all seven domains: passed with zero warnings, zero assertion failures, zero failed stages, and exact Queue reconciliation; 15 typed Queue 4005 retries recovered; Stream max latency 6.37s under the 10s progress gate

AI-assisted: yes