…LM keep-alive, fix typo)

…ions

- Add SetAnalysis(key, analysis) to correlation engine for write-back
- Wire analysisWriter callback from AlertManager to engine
- Bump enrichCh buffer 1→8 to reduce dropped enrichment under burst
- Write analysis back to original incident so subsequent updates include it
- Swap LLM model from Qwen2.5-1.5B Q4_K_M to MiniCPM5-1B Q4_K_M (688 MB)
- Add --jinja flag, adjust temp 0.1→0.7, top-p 0.9→0.95, drop top-k
- Reduce LLM sidecar resources: cpu 1→500m, memory 2Gi→1Gi
- Remove deprecated Modelfile (content moved to code)
- Add linux/arm/v7 and linux/arm/v6 to kwatch publish workflow

- buildSeenSet: capture Running containers with RestartCount > 0 using
  LastTerminationState.Terminated.Reason so CrashLoopBackOff pods are
  baselined even when briefly in Running state between crashes
- SetSeen: merge (not replace) the seen set so entries from persistent
  ConfigMap baseline survive when buildSeenSet misses some pods; remove
  !exists guard so timestamps get refreshed by buildSeenSet scans

- deliverOne: accept context so sendWithRetry respects shutdown
  context, preventing long retries from blocking graceful shutdown
- startBaselineSaver: stop timer before Reset to avoid Go anti-pattern
- baseline channel: increase buffer to 64 and drop drain pattern,
  eliminating the TOCTOU blocking risk between concurrent callers
- sleepWithContext: remove unreachable nil-guard
- engine: add 'Caller must hold e.mu.' comments on 8 internal
  functions that rely on caller holding the mutex

- Hardcode replicas: 1 (kwatch must always run as a single replica)
- Remove nativeSidecar toggle (always use plain container, never initContainer)
- Remove both replica guard fail blocks
- Clean up tests and README accordingly
- Add upgrader.disableUpdateCheck row to chart README

- Chart.yaml: Qwen2.5-1.5B via Ollama -> MiniCPM5-1B via llama.cpp
- config.go: localhost:11434 -> localhost:8080
- alert_test.go: Stub Ollama -> Stub LLM sidecar
- Remove CHANGELOG.md

- New SVG logo: hexagonal badge with radar scope, concentric rings,
  8 resource blips, sweep arc, and bold K letterform
- Brand gradient from K8s blue (#326CE5) to kwatch cyan (#00ACD7)
- Update README hero text to reflect accurate capabilities

Co-authored-by: AI <ai@kwatch.dev>

- Fix publish-llm.yml image name to kwatch-llm
- Update LLM image tags to dev/latest
- Bump LLM VERSION to dev

Increased MaxBaseline and adjusted ResolveHoldDown value.

Updated the system prompt for Kubernetes RCA assistant to clarify instructions and improve readability.

… fix TestFirstKey flake; periodic graph rebuild

…er nil guards, AddIndexers unchecked error, shutdown goroutine lifecycle, metrics dead code

…ppression

- Add custom enricher mock to test unreachable PodHasIssues return path
- Add DetectControlPlanePodIssue waiting fallback (ImagePullBackOff) coverage
- Add emitHighRestartAlert empty owner early return test
- Add buildContainerHint LivenessProbeFailed and OOMRepeating tests
- Add event lister multi-event sort path tests for both pod and container filters
- Add unschedulable delay fallback (delay <= 0) test

…graph pruning, service endpoint debounce, LLM prompt fix, and various improvements

- Add jitter support to alert provider retry with configurable jitterFactor
- Add ClusterAutoscalerMonitor for cluster-autoscaler event alerting
- Add AuditLog config for structured JSON audit logging of incidents
- Add graph pruning (mark-and-sweep) to prevent stale dependency entries
- Fix LLM system prompt to focus only on logs, never mention infra names
- Add 60s sustained debounce for service endpoint notifications
- Fix smart grouping scope: pod-level reasons no longer scoped as node
- Add ServiceNoEndpoints to obviousReasons to skip LLM enrichment
- Fix re-notification guard in smart grouping (NotifiedSig check)
- Add RBAC configmaps permission for readiness probe fix
- Add OOM timeline reporting to container hints
- Add dependency-filtered change detection in insight engine
- Add cross-namespace ImagePullBackOff dedup for global scope issues
- Fix SnapshotAll dirty flag optimization
- Update README with corrected defaults and debounce documentation

- Conditional fields (skip empty values) instead of rigid templates
- Natural language format with middle dot (·) separators, no pipes
- Event/Logs shown only when IncludeEvents/IncludeLogs + non-empty
- Extracted resourcePlural() helper to eliminate repeated plural logic
- Extracted thread map helpers (saveThread/loadThread/popThread) in Slack
- All providers updated: core, Slack, Discord, Mattermost, Opsgenie,
  Teams, Telegram, Email, Zenduty, PagerDuty, Webhook, event formatters
- go test ./... PASS