Fix dependency cycles, cost budgets, scaling backtesting, prefetch limits

[Realericky]

Summary

Addresses four independent issues across the observability/performance modules:

• backend/src/dependency_map.rs — POST /dependencies/discover now rejects an edge that would introduce a cycle (a direct self-loop, or a path that already exists from to back to from), via new DependencyGraph::has_path / would_create_cycle helpers. Rejected with 422 and a clear error message. The existing DOT export (GET /dependencies/graph) already serves as the graph's visualization output, documented accordingly.
• backend/src/cost_tracking.rs — adds configurable per-category BudgetThresholds, scoped either by operation or by a tag key/value pair (covering per-vault gas and per-tenant API cost budgets). POST /admin/cost/entries re-evaluates thresholds on every recorded entry and logs a warning on breach. New POST /admin/cost/budget-thresholds and GET /admin/cost/budget-breaches endpoints, plus breach_to_incident_request / breach_to_escalation_request helpers that convert a breach into an incidents::CreateIncidentRequest / oncall::TriggerEscalationRequest, wiring budget alerts into those workflows.
• backend/src/predictive_scaling.rs — adds ForecastModel::backtest (and a PredictiveScaler::backtest convenience wrapper) that replays historical TrafficSample data, forecasting ahead at each point with sufficient history and comparing against the actual sample that later arrived, returning MAE/MAPE/RMSE (BacktestResult) so model accuracy can be validated before it drives live scaling decisions.
• backend/src/cache_warming.rs — CacheWarmer now caps concurrent prefetch execution (max_concurrent_prefetches, default 10, configurable via CacheWarmer::with_concurrency_limit), executing predict_prefetch_targets in capped chunks via futures::future::join_all so a large candidate set can't spike load on the origin store at once — the existing confidence-descending sort means the highest-priority targets always get through first when capped. Also adds a prefetch hit-rate-vs-cost metric (WarmerStats::prefetch_hit_rate, exported via CacheWarmer::render_prometheus).

Additionally, cache_warming, dependency_map, incidents and oncall were not declared in backend/src/lib.rs's module tree at all (no pub mod entry), so none of their code — including the fixes above — was actually part of the compiled crate. Added the missing pub mod declarations so these modules (and the cost_tracking → incidents/oncall wiring) are reachable.

Each corresponding doc (docs/dependency-mapping.md, docs/cost-tracking.md, docs/predictive-scaling.md, docs/cache-warming.md) was updated to describe the new behavior.

Test plan

• cargo build -p ethos-protocol-backend / cargo test -p ethos-protocol-backend (not run in this environment — no Rust toolchain available)
• Manually exercise POST /dependencies/discover with an edge that closes a cycle and confirm 422
• Manually configure a budget threshold and confirm GET /admin/cost/budget-breaches reports it once crossed
• Manually call PredictiveScaler::backtest() / ForecastModel::backtest() against recorded traffic history
• Manually run CacheWarmer::warm_cache against a candidate set larger than the concurrency cap and confirm chunked execution

Closes #378
Closes #379
Closes #380
Closes #381

[drips-wave]

@Realericky Great news! 🎉 Based on an automated assessment of this PR, the linked Wave issue(s) no longer count against your application limits.

You can now already apply to more issues while waiting for a review of this PR. Keep up the great work! 🚀

Learn more about application limits