docs(claude): service map is the sunflower layout + pre-sample topology observer (#118)

Update the project instructions to match the code after #117:
- The service map uses the deterministic phyllotaxis ("sunflower") layout
  (radialLayout.ts), not the (now-removed) dagre layered layout; edges are
  border-pinned chords with direction arrowheads; selection accents the
  active path (nodes + edges). @dagrejs/dagre is removed.
- Document the pre-sample topology observer (topology_observer.go) that
  records cross-service CALLS edges independent of SAMPLING_RATE, why it
  exists (joint-survival starvation at low sampling), and its existence-only
  + bounded design so sampled aggregates and OOM bounds are unaffected.


Claude-Session: https://claude.ai/code/session_01LDQVJrixs2nJoea67a8pEG

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

Author: aksOps

CLAUDE.md

@@ -4,7 +4,7 @@
 
 OtelContext is a self-hosted OTLP observability platform. Single Go binary with embedded React frontend.
 - **Backend:** Go 1.25, native `net/http` (no frameworks), GORM ORM, gRPC + HTTP for OTLP ingestion
-- **Frontend:** React 19 + TypeScript + TanStack Query/Virtual + wouter + Radix primitives + cmdk palette + hand-rolled token CSS (`ui/src/styles/tokens.css`) with CSS Modules. The **service map renders via React Flow (`@xyflow/react`, MIT)** over a dagre layered layout (`ui/src/lib/dagreLayout.ts` — `@dagrejs/dagre`, MIT: crossing-minimised, left→right, with edge-routing waypoints) computed synchronously. React Flow owns pan/zoom/fit/minimap/grid/a11y; nodes are token-themed React components — draggable, and they collapse to status **dots** below a zoom threshold (level-of-detail) so they stop occluding edges at scale; edges route through the dagre waypoints and meet node borders. No UI framework: `@ossrandom/design-system`, cytoscape, uplot remain removed (cytoscape's physics hairball → deterministic layered layout + React Flow; the prior hand-rolled-SVG map was replaced by React Flow on 2026-06-18, then moved to the dagre layered layout)
+- **Frontend:** React 19 + TypeScript + TanStack Query/Virtual + wouter + Radix primitives + cmdk palette + hand-rolled token CSS (`ui/src/styles/tokens.css`) with CSS Modules. The **service map renders via React Flow (`@xyflow/react`, MIT)** over a deterministic phyllotaxis ("sunflower") layout (`ui/src/lib/radialLayout.ts` — most-critical service near the disc centre, golden-angle spiral) computed synchronously. Unlike a layered DAG, the sunflower fills a disc evenly and never collapses sparse/disconnected graphs into a single column, so the map stays navigable from 7 to 120+ services. React Flow owns pan/zoom/fit/minimap/grid/a11y; nodes are token-themed React components — draggable, and they collapse to status **dots** below a zoom threshold (level-of-detail) so they stop occluding edges at scale. Edges are straight chords pinned to node borders (centre-to-centre at dot zoom) carrying a direction **arrowhead** (`markerEnd`) so flow is legible statically — at any zoom and under `prefers-reduced-motion`; selecting a node accents the active path (both nodes **and** edges) and dims the rest. No UI framework: `@ossrandom/design-system`, cytoscape, uplot, **and `@dagrejs/dagre`** remain removed (cytoscape's physics hairball → React Flow; the prior hand-rolled-SVG map was replaced by React Flow on 2026-06-18; the dagre layered layout it briefly used was itself replaced by the sunflower layout on 2026-06-18 because a layered DAG stacked sparse 120-service graphs into an unreadable single column)
 - **Ports:** gRPC `:4317` (OTLP), HTTP `:8080` (API + HTTP OTLP + WebSocket + UI)
 
 ## Strict Rules
@@ -114,6 +114,8 @@ LogsServer.Export()  → DB persist → logCallback  → GraphRAG.OnLogIngested(
 MetricsServer.Export() → TSDB    → metricCallback → GraphRAG.OnMetricIngested()
 ```
 
+**Pre-sample topology observer** (`internal/graphrag/topology_observer.go`): `TraceServer.Export()` also calls `GraphRAG.ObserveSpanTopology()` for **every** received span *before* the sampler's keep/drop decision, so cross-service CALLS edges are recorded independent of `SAMPLING_RATE`. Without it, an edge formed only when both the caller and callee spans survived sampling (~`rate²` joint probability), so at low sampling the service map showed nodes but almost no edges/flow. The observer is existence-only (`ServiceStore.EnsureService`/`EnsureCallEdge` create with zeroed aggregates, no-op if present) so the sampled path stays the **sole** source of CallCount/latency/error-rate; it's bounded by a per-tenant spanID→service LRU (cap 100k) + per-pair dedup (memory ≈ #service-pairs) and never touches TraceStore/eventCh, so the SQLite OOM-survival bounds are unaffected.
+
 ## MCP Server — 7-Tool Triage Surface
 
 The MCP server (`internal/mcp/`) exposes a focused 7-tool triage surface via