diff --git a/docs/specs/resolved-path.md b/docs/specs/resolved-path.md
index 61e7286..6d316c0 100644
--- a/docs/specs/resolved-path.md
+++ b/docs/specs/resolved-path.md
@@ -17,21 +17,22 @@ Currently, hop paths are stored as short uppercase hex prefixes in `path_json` (
 
 ## Solution
 
-Resolve hop prefixes to full pubkeys **once at ingest time** on the server, using `resolveWithContext()` with 4-tier priority (affinity → geo → GPS → first match) and a **persisted neighbor graph**. Store the result as a new `resolved_path` field alongside `path_json`.
+Resolve hop prefixes to full pubkeys **once at ingest time** on the server, using `resolveWithContext()` with 4-tier priority (affinity → geo → GPS → first match) and a **persisted neighbor graph**. Store the result as a new `resolved_path` column on observations alongside `path_json`.
 
 ## Design decisions (locked)
 
 1. **`path_json` stays unchanged** — raw firmware prefixes, uppercase hex. Ground truth.
-2. **`resolved_path` is NEW** — `[]*string` on observations. Full 64-char lowercase hex pubkeys, `null` for unresolved.
+2. **`resolved_path` is a column on observations** — full 64-char lowercase hex pubkeys, `null` for unresolved.
 3. **Resolved at ingest** using `resolveWithContext(hop, context, graph)` — 4-tier priority: affinity → geo → GPS → first match.
-4. **Stored alongside `path_json`** — raw firmware prefixes stay unchanged.
-5. **`null` = unresolved** — ambiguous prefixes store `null`. Frontend falls back to prefix display.
-6. **Both fields coexist** — not interchangeable. Different consumers use different fields.
+4. **`null` = unresolved** — ambiguous prefixes store `null`. Frontend falls back to prefix display.
+5. **Both fields coexist** — not interchangeable. Different consumers use different fields.
 
 ## Persisted neighbor graph
 
 ### SQLite table: `neighbor_edges`
 
+Thin and normalized. Stores ONLY the relationship. SNR, observer names, GPS, roles — all join from existing tables when needed. No duplication.
+
 ```sql
 CREATE TABLE IF NOT EXISTS neighbor_edges (
     node_a TEXT NOT NULL,
@@ -42,24 +43,42 @@ CREATE TABLE IF NOT EXISTS neighbor_edges (
 );
 ```
 
-### Order of operations
+### Edge extraction rules (ADVERT vs non-ADVERT)
 
-1. **First run:** Built on startup from existing packet data, then persisted to SQLite.
-2. **Subsequent runs:** Loaded from SQLite — instant startup, no 7-second rebuild.
-3. **Available at ingest** — lives on `PacketStore`, not `Server`.
-4. **Updated incrementally** — each new packet upserts edges.
+At ingest, for each packet:
 
-This replaces the previous "in-memory only for v1" approach. The graph is always persisted, always fast to load.
+- **ADVERT packets** (payload_type 4): originator pubkey is known from `decoded_json.pubKey`. Extract edge: `originator ↔ path[0]` (the first hop is a direct neighbor of the originator).
+- **ALL packets**: observer pubkey is known. Extract edge: `observer ↔ path[last]` (the last hop is a direct neighbor of the observer).
+- **Non-ADVERT packets**: originator is unknown (encrypted). ONLY extract `observer ↔ path[last]`.
+- Each packet produces **1 or 2 edge upserts** depending on type.
+
+Edge upsert: `INSERT OR REPLACE INTO neighbor_edges` with `count = count + 1` and `last_seen = now`.
+
+### Cold startup and backfill
+
+On startup:
+
+1. **Load `neighbor_edges` from SQLite** → build in-memory graph.
+2. **If table empty (first run):** `BuildFromStore(packets)` — scan all existing packets, extract edges per the rules above, INSERT into `neighbor_edges`.
+3. **Load observations from SQLite.**
+4. **For observations without `resolved_path`:** resolve using the graph, UPDATE `resolved_path` in SQLite.
+5. **Ready to serve.**
+
+On subsequent runs, step 2 is skipped (table already populated). Step 4 only processes observations with NULL `resolved_path` (new or previously unresolved).
 
 ## Data model
 
 ### Where does `resolved_path` live?
 
-**On observations**, not transmissions.
+**On observations**, as a column:
+
+```sql
+ALTER TABLE observations ADD COLUMN resolved_path TEXT;
+```
 
 Rationale: Each observer sees the packet from a different vantage point. The same 2-char prefix may resolve to different full pubkeys depending on which observer's neighborhood is considered. The observer's own pubkey provides critical context for `resolveWithContext` (tier 2: neighbor affinity). Storing on observations preserves this per-observer resolution.
 
-`path_json` already lives on observations in the DB schema. `resolved_path` follows the same pattern.
+`resolved_path` is written in the same INSERT that creates the observation — one write, no double-write problem.
 
 ### Field shape
 
@@ -72,18 +91,37 @@ resolved_path TEXT  -- JSON array: ["aabb...64chars", null, "ccdd...64chars"]
 - Stored as a JSON text column (same approach as `path_json`)
 - Uses `omitempty` — absent from JSON when not set
 
+## Every path resolution uses the graph — no exceptions
+
+All existing `pm.resolve()` call sites MUST be migrated to `resolveWithContext` with the persisted graph. No "we'll get to it later."
+
+### Call sites to migrate (exhaustive)
+
+Found via `grep -n "pm.resolve" cmd/server/store.go`:
+
+| Line | Function | Current | After |
+|------|----------|---------|-------|
+| 1192 | `IngestNewFromDB()` | `pm.resolve(hop)` | `resolveWithContext(hop, ctx, graph)` — resolve at ingest, store as `resolved_path` |
+| 1876 | `buildDistanceIndex()` | `pm.resolve(hop)` | Read `resolved_path` from observation — already resolved at ingest |
+| 3537 | `computeAnalyticsTopology()` | `pm.resolve(hop)` | Read `resolved_path` from observation |
+| 5528 | `computeAnalyticsSubpaths()` | `pm.resolve(hop)` | Read `resolved_path` from observation |
+| 5665 | `GetSubpathDetail()` | `pm.resolve(hop)` | `resolveWithContext(hop, ctx, graph)` — ad-hoc resolution for user-provided hops |
+| 5744 | `GetSubpathDetail()` | `pm.resolve(h)` | `resolveWithContext(h, ctx, graph)` — same function, second usage |
+
+**After migration:** `pm.resolve()` (naive prefix-only lookup) is dead code. Remove it. All resolution goes through `resolveWithContext` which uses the persisted neighbor graph for affinity-based disambiguation.
+
 ## Ingest pipeline changes
 
 ### Where resolution happens
 
-In `PacketStore.IngestNewFromDB()` and `PacketStore.IngestNewObservations()` in `cmd/server/store.go`. Resolution is added **after** the observation is read from DB and **before** it's stored in memory and broadcast.
+In `PacketStore.IngestNewFromDB()` in `cmd/server/store.go`. Resolution is added **during** the observation INSERT — same write, same transaction.
 
 Resolution flow per observation:
 1. Parse `path_json` into hop prefixes
 2. Build context pubkeys from the observation (observer pubkey, source/dest from decoded packet)
 3. Call `resolveWithContext(hop, contextPubkeys, neighborGraph)` for each hop
-4. Store result as `resolved_path` on the observation
-5. Upsert neighbor edges into the graph (incremental update)
+4. Store result as `resolved_path` column on the observation (same INSERT)
+5. Upsert neighbor edges into `neighbor_edges` table (incremental update)
 
 ### Performance
 
@@ -102,8 +140,10 @@ Per-hop cost: ~1–5μs. A typical packet has 0–5 hops. At 100 packets/second
 | Map Show Route | Client HopResolver (naive) | Read `resolved_path` |
 | Live map animated paths | Client HopResolver (naive) | Read `resolved_path` |
 | Node detail paths | Client HopResolver (naive) | Read `resolved_path` |
-| Analytics topology/subpaths | Server `prefixMap.resolve()` (naive) | Read `resolved_path` from observations |
-| Analytics hop distances | Server `resolveWithContext` | Already correct |
+| Analytics topology | Server `pm.resolve()` (naive) | Read `resolved_path` from observations |
+| Analytics subpaths | Server `pm.resolve()` (naive) | Read `resolved_path` from observations |
+| Analytics hop distances | Server `pm.resolve()` (naive) | Read `resolved_path` from observations |
+| Subpath detail | Server `pm.resolve()` (naive) | `resolveWithContext` with graph |
 | Show Neighbors | Server neighbors API | Already correct |
 | `/api/resolve-hops` | Server `resolveWithContext` | Already correct |
 | Hex breakdown display | `path_json` raw | Unchanged — shows raw bytes |
@@ -133,11 +173,10 @@ All endpoints that currently return `path_json` also return `resolved_path`:
 - **`path_json` display prefixes:** uppercase (raw firmware)
 - **`resolved_path`:** lowercase full pubkeys
 - **Comparison code:** normalizes to lowercase
-- **No case standardization crusade** — document the convention, don't change existing behavior
 
 ## Backward compatibility
 
-- Old observations without `resolved_path`: frontend falls back to client-side HopResolver (same as today).
+- Old observations without `resolved_path`: resolved during cold startup backfill (step 4). If still `null` after backfill, frontend falls back to client-side HopResolver.
 - `resolved_path` field uses `omitempty` — absent from JSON when not set.
 
 ### Fallback pattern (frontend)
@@ -153,19 +192,21 @@ function getResolvedHops(packet) {
 ## Implementation milestones
 
 ### M1: Persist graph to SQLite + load on startup + incremental updates at ingest
-- Create `neighbor_edges` table in SQLite
-- Build graph from existing packet data on first run
-- Load from SQLite on subsequent runs (instant startup)
+- Create `neighbor_edges` table in SQLite (schema above)
+- On first run: `BuildFromStore(packets)` — scan all packets, extract edges per ADVERT/non-ADVERT rules, INSERT into table
+- On subsequent runs: load from SQLite → build in-memory graph (instant startup)
 - Upsert edges incrementally during packet ingest
 - Graph lives on `PacketStore`, not `Server`
-- Tests: graph persistence, load, incremental update
+- Tests: graph persistence, load, incremental update, ADVERT vs non-ADVERT edge extraction
 
-### M2: Add `resolved_path` to observations at ingest time
+### M2: Add `resolved_path` column to observations + resolve at ingest
+- `ALTER TABLE observations ADD COLUMN resolved_path TEXT`
 - Add `ResolvedPath []*string` to `Observation` struct
-- Resolve during `IngestNewFromDB`/`IngestNewObservations`
-- Write `resolved_path` back to SQLite observations table
-- Schema migration to add `resolved_path` column
-- Tests: unit test resolution at ingest, verify stored values
+- Resolve during `IngestNewFromDB` — same INSERT, one write
+- Cold startup backfill: resolve observations with NULL `resolved_path`, UPDATE in SQLite
+- Migrate ALL 6 `pm.resolve()` call sites to `resolveWithContext` or read from `resolved_path`
+- Remove dead `pm.resolve()` code
+- Tests: unit test resolution at ingest, verify stored values, verify all call sites use graph
 
 ### M3: Update all API responses to include `resolved_path`
 - Include `resolved_path` in all packet/observation API responses
@@ -173,7 +214,7 @@ function getResolvedHops(packet) {
 - Tests: verify API response shape, WS broadcast shape
 
 ### M4: Update frontend consumers to prefer `resolved_path`
-- Update `packets.js`, `map.js`, `live.js`, `analytics.js`, `nodes.js`, `traces.js`
+- Update `packets.js`, `map.js`, `live.js`, `analytics.js`, `nodes.js`
 - Add fallback to `path_json` + `HopResolver` for old packets
 - `hop-resolver.js` becomes fallback only
 - Tests: Playwright tests for path display