Implement duplicate suppression and noise-budget engine

[mdheller]

Goal

Implement event coalescing and noise budgeting so repeated low-level records do not become operator-facing garbage.

Scope

• Define event fingerprint algorithm.
• Add duplicate suppression window.
• Add per-daemon/component noise budget fields.
• Track raw-to-canonical event reduction ratio.
• Preserve representative evidence samples.
• Escalate only when repeated events change risk class, not because repetition is noisy.

Acceptance criteria

• Repeated identical denials collapse into one canonical telemetry.coalesced event.
• Coalesced events preserve count, first_seen, last_seen, and sample evidence references.
• Severity is not inflated by duplication alone.
• DeliveryExcellence can consume signal-quality metrics.

[mdheller]

Status (2026-07-18): Duplicate suppression and noise-budget engine is not yet implemented. The event coalescing logic (fingerprint algorithm, dedup window, per-daemon noise budgets, raw-to-canonical reduction ratio, representative sample preservation) is deferred to a future sprint. The canonical event store from PR #42 provides the foundation for this work. Keeping open.

[mdheller]

Implemented in src/sourceos_syncd/noise_budget.py (PR #51). NoiseBudgetEngine does per-key SHA256 dedup with configurable window; critical events (state-integrity-violation, policy-violation, repair-required, conflict-detected, identity-loss) bypass the budget and always surface. CLI: sourceos-syncd noise-budget status and noise-budget flush. Service endpoint: /noisez.