Jevops: SRE log triage with Jev

# Jevops

SRE log tool built on **TypeSafe Jev** (typed decision model) with Elasticsearch, Logstash, Kibana, and Grafana integration — two coexisting approaches:

- **Mode A — real-time triage:** streams logs and decides page / digest / resolve (see below).
- **Mode B — investigative query:** answers *"what's important from 8am to now?"* by hierarchically drilling chunks of logs with Jev, then having an LLM explain the exact lines with citations and tool calls (`docs/hierarchical-query.md`).

Jev answers typed questions against a state (`choice` probabilities, `score` levels, `noul` 0–1 likelihoods) with calibrated confidence, ~70–500 ms, $0.042/M input tokens (output free). Jevops turns that into pages, digests, and resolved incidents instead of log-sprawl — or into a cited timeline when you ask a question.

```
                    ┌──────────────┐   HTTP /ingest
app logs ──► Logstash├─► Elasticsearch ──► poll-es ──┐
      (tcp :5000)   └─► Jevops intake ───────────────┤
                                                     ▼
                       prefilter (free) ──► burst rank (1 Jev call / window)
                                                     ▼
                       triage fan-out (severity, category, urgency, needs_human,
                                        novelty, recovery — all parallel)
                                                     ▼
                       policy (confidence floor, cooldown, composite score)
                          ├─► page ──► PagerDuty / Slack / webhook
                          ├─► resolve ──► PagerDuty (dedup_key)
                          └─► digest ──► SQLite + Elasticsearch ◄── Grafana dashboard
```

## Why Jev here

- **Burst ranking:** up to 255 log lines in one request (`choice` over line IDs + `page_worthy` `nou