diff --git a/CLAUDE.md b/CLAUDE.md
new file mode 100644
index 00000000..9054813d
--- /dev/null
+++ b/CLAUDE.md
@@ -0,0 +1,352 @@
+# Golbat Architecture Guide
+
+## Overview
+
+Golbat is a high-performance Go backend that receives raw protobuf data from Pokemon GO game clients, decodes it, maintains an in-memory cache of all game entities, persists changes to MySQL via write-behind queues, dispatches webhooks, and serves a REST/gRPC API for querying entities with spatial and attribute-based filters.
+
+## Project Layout
+
+```
+main.go              — HTTP/gRPC server setup, route registration
+routes.go            — HTTP route handlers (raw ingest, API endpoints)
+decode.go            — Proto method dispatcher, GMO decoder
+grpc_server_raw.go   — gRPC raw proto receiver
+decoder/
+  main.go            — Cache/queue initialization, raw data types
+  sharded_cache.go   — Generic sharded TTL cache
+  scanarea.go        — ScanParameters and scan rule matching
+  gmo_decode.go      — Batch processors (forts, pokemon, weather, stations)
+  <entity>.go        — Struct definition, setters with dirty tracking
+  <entity>_state.go  — CRUD: load/save/webhooks, get*Record* functions
+  <entity>_decode.go — Proto → entity field mapping
+  <entity>_process.go — High-level proto processing (FortDetails, encounters, etc.)
+  api_<entity>.go    — API result structs, scan endpoints, DNF filters
+  pokemonRtree.go    — Pokemon spatial index + lookup cache
+  fortRtree.go       — Fort spatial index + lookup cache
+  fort_tracker.go    — In-memory fort lifecycle tracking via S2 cells
+  tracked_mutex.go   — Lock contention instrumentation
+  writebehind/       — Write-behind queue implementation
+  writebehind_batch.go — Queue initialization and batch flush SQL
+webhooks/
+  webhook.go         — Webhook types, config parsing, HTTP dispatch
+  sender.go          — Thread-safe message batching and sending
+config/              — TOML config parsing
+db/                  — Database connection details, query helpers
+geo/                 — Geofence loading, R-tree matching, S2 lookup
+```
+
+## Raw Message Processing
+
+### Ingest Endpoints
+
+Messages arrive via two paths:
+
+1. **HTTP POST `/raw`** (`routes.go`): Accepts JSON with base64-encoded protobuf payloads. Supports both Pogodroid format (array of `{payload, type}`) and standard format (object with `contents[]`, `username`, `trainerlvl`, etc.). Returns 201 immediately; processing happens in a background goroutine with a 5s timeout.
+
+2. **gRPC `SubmitRawProto`** (`grpc_server_raw.go`): Accepts `RawProtoRequest` with binary proto payloads in `Contents[]`. Same async processing pattern.
+
+Both paths normalize into `ProtoData` structs and call `decode()`.
+
+### Dispatch (`decode.go`)
+
+`decode()` switches on `pogo.Method` to route each proto type:
+
+- `METHOD_GET_MAP_OBJECTS` → `decodeGMO()` — the primary data source
+- `METHOD_FORT_DETAILS` → individual fort/gym detail updates
+- `METHOD_GYM_GET_INFO` → gym defender/detail updates
+- `METHOD_ENCOUNTER` / `METHOD_DISK_ENCOUNTER` → pokemon encounter data
+- `METHOD_FORT_SEARCH` → quest rewards
+- `METHOD_GET_MAP_FORTS` → bulk fort name/image data
+- Plus ~15 other method types (invasions, routes, tappables, weather, etc.)
+
+Level 30+ is required for most methods to ensure data quality.
+
+### GetMapObjects (GMO) Processing
+
+`decodeGMO()` is the main data pipeline. It parses `GetMapObjectsOutProto` and extracts:
+
+- **Forts** (pokestops + gyms) → `UpdateFortBatch()`
+- **Wild/Nearby/Map Pokemon** → `UpdatePokemonBatch()`
+- **Weather** → `UpdateClientWeatherBatch()` → triggers proactive IV switching
+- **Stations** → `UpdateStationBatch()`
+- **S2 Cells** → cell timestamp tracking
+- **Fort removal detection** → `CheckRemovedForts()`
+
+Processing is gated by `ScanParameters` — a set of boolean flags resolved from config scan rules based on the request's `scan_context` and geographic location. This allows different scan areas to process different entity types.
+
+## Entity Model
+
+### Struct Pattern
+
+Every entity follows the same pattern:
+
+```go
+type PokestopData struct {    // Copyable data fields with db tags
+    Id   string  `db:"id"`
+    Name null.String `db:"name"`
+    // ... all persisted columns
+}
+
+type Pokestop struct {
+    mu TrackedMutex[string] `db:"-"`  // Entity-level mutex
+    PokestopData                       // Embedded — copied for queue snapshots
+    dirty     bool     `db:"-"`        // Needs DB write
+    newRecord bool     `db:"-"`        // INSERT vs UPDATE
+    oldValues PokestopOldValues `db:"-"` // Snapshot for webhook comparison
+}
+```
+
+- **Setter methods** (`SetName`, `SetLat`, etc.) track dirty state and optionally log field changes when `dbDebugEnabled` is true.
+- **`snapshotOldValues()`** captures current field values before modifications, used later for webhook change detection.
+- The embedded `Data` struct can be copied by value for the write-behind queue without copying the mutex or internal state.
+
+### Entities
+
+| Entity | Cache Key | Cache Type | Queue Key | ID Type |
+|--------|-----------|------------|-----------|---------|
+| Pokestop | string (fort ID) | Sharded | string | string |
+| Gym | string (fort ID) | Sharded | string | string |
+| Pokemon | uint64 (encounter ID) | Sharded | uint64 | uint64 |
+| Station | string (station ID) | Sharded | string | string |
+| Spawnpoint | int64 (spawn ID) | Sharded | int64 | int64 |
+| Incident | string (incident ID) | TTL | string | string |
+| Weather | int64 (S2 cell ID) | TTL | — | int64 |
+| Route | string (route ID) | TTL | string | string |
+| Tappable | uint64 (encounter ID) | TTL | uint64 | uint64 |
+
+## Caching
+
+### ShardedCache
+
+High-contention entities (pokestop, gym, station, spawnpoint, pokemon) use `ShardedCache[K, V]` — a generic wrapper over multiple `ttlcache.Cache` instances. Keys are distributed across `runtime.NumCPU()` shards via FNV-1a hashing (strings) or identity (integers), reducing lock contention on the underlying cache maps.
+
+### TTL Cache
+
+Lower-contention entities (incident, weather, route, tappable, player, s2cell) use a single `ttlcache.Cache` instance.
+
+### Configuration
+
+- **Fort caches**: 60-minute TTL normally, 25 hours when `config.Config.FortInMemory` is enabled (keeps forts resident for R-tree operations).
+- **Pokemon cache**: 60-minute TTL with `DisableTouchOnHit = true` — TTL counts from creation, not last access, ensuring pokemon expire after their despawn time regardless of query frequency.
+- **All other caches**: 60-minute TTL (weather consensus: 2 hours).
+
+### Eviction Callbacks
+
+Fort and pokemon caches register eviction callbacks that clean up the corresponding R-tree entries when a cache item expires. This maintains consistency between the cache and spatial index.
+
+## Locking Model
+
+### Entity-Level Mutex
+
+Each entity instance has its own mutex (`TrackedMutex`). All access to entity fields goes through `get*Record*` functions that return `(entity, unlockFunc, error)`. The caller MUST call the returned unlock function.
+
+### Record Access Patterns
+
+Four access patterns, from lightest to heaviest:
+
+1. **`Peek*Record`** — Cache-only lookup, no DB fallback. Returns locked entity or nil. Used for read-only API queries where missing data is acceptable.
+
+2. **`get*RecordReadOnly`** — Cache lookup with DB fallback on miss. Acquires lock but does NOT snapshot old values. Used for read-only access that needs complete data.
+
+3. **`get*RecordForUpdate`** — Calls ReadOnly internally, then snapshots old values for webhook comparison. Used when modifying an existing entity.
+
+4. **`getOrCreate*Record`** — Atomically creates a new cache entry if absent (via `GetOrSetFunc`), then locks and loads from DB if marked as new record. Always snapshots. Used when the entity may not exist yet.
+
+### Atomic Cache Population
+
+`GetOrSetFunc` ensures only one goroutine creates a given cache entry. If two goroutines race to create the same entity, one wins and the other gets the winner's instance. Both then lock the same mutex, serializing their updates.
+
+### Lock Ordering
+
+**Never hold two entity locks simultaneously.** When multiple entities must be accessed (e.g., pokestop-to-gym conversion copying shared fields), release the first lock before acquiring the second. The pattern is: lock A → copy needed data → unlock A → lock B → apply data → unlock B.
+
+If you must reason about lock priority (e.g., choosing which to acquire first), use this ordering by dependency:
+
+```
+1. Pokestop / Gym      (peers — never lock both at once)
+2. Station
+3. Incident             (references Pokestop for lat/lon/name)
+4. Pokemon              (references Pokestop, Spawnpoint)
+5. Spawnpoint
+6. Weather
+7. Route
+8. Tappable
+```
+
+In practice most code only locks a single entity. The cases where two interact:
+- **Incident/Pokemon save** → briefly locks Pokestop to copy lat/lon/name, then releases
+- **Fort type conversion** → copies shared fields from one fort type to the other with release-between
+
+## Write-Behind Queues
+
+### Architecture
+
+Each entity type has a `TypedQueue[K, T]` that batches and coalesces writes:
+
+1. **Enqueue**: Takes a snapshot of the entity's Data struct (value copy) and adds it to a pending map keyed by entity ID. If the same entity is enqueued again before flushing, the entry is updated in-place (coalescing).
+
+2. **Dispatch**: A processing loop checks for entries whose `ReadyAt` time has passed, moves them to a batch buffer.
+
+3. **Flush**: When the batch reaches `BatchSize` (default 50) or `BatchTimeout` (default 100ms) elapses, the batch is flushed via a bulk `INSERT ... ON DUPLICATE KEY UPDATE` SQL statement.
+
+4. **Concurrency**: All queues share a `SharedLimiter` that caps total concurrent DB writers (default 50). This prevents overwhelming the database connection pool.
+
+5. **Deadlock retry**: MySQL deadlock errors (1213) trigger up to 3 retries with exponential backoff.
+
+### Pokemon Delay
+
+Wild and nearby pokemon use a 30-second write delay (`wildPokemonDelay`). When a wild pokemon is first seen in a GMO, it's enqueued with `delay = 30s`, giving time for an encounter request to arrive with IV/CP/level data. If an encounter arrives within the window, the queue entry is updated in-place with the richer data. Encounter-sourced pokemon write immediately (delay = 0).
+
+### Database
+
+**Golbat targets MariaDB.** MySQL may work but is not tested. SQL syntax, migrations, and batch upsert queries are written for MariaDB compatibility.
+
+#### Connection Split
+
+`DbDetails` holds two connection pools:
+- **`PokemonDb`**: Dedicated pool for the pokemon table (highest write volume).
+- **`GeneralDb`**: Everything else (forts, incidents, weather, routes, etc.).
+
+## Webhooks
+
+### Types
+
+| Config String | Webhook Type | Payload Type |
+|---------------|-------------|--------------|
+| `pokemon_iv` | PokemonIV | `pokemon` |
+| `pokemon_no_iv` | PokemonNoIV | `pokemon` |
+| `pokemon` | Both IV types | `pokemon` |
+| `gym` | GymDetails | `gym_details` |
+| `raid` | Raid | `raid` |
+| `quest` | Quest | `quest` |
+| `pokestop` | Pokestop | `pokestop` |
+| `invasion` | Invasion | `invasion` |
+| `weather` | Weather | `weather` |
+| `fort_update` | FortUpdate | `fort_update` |
+| `max_battle` | MaxBattle | `max_battle` |
+
+### Dispatch Flow
+
+1. After saving an entity, the save function calls webhook creation functions (e.g., `createPokemonWebhooks`, `createGymFortWebhooks`).
+2. These build a webhook payload struct and call `webhooksSender.AddMessage(type, payload, areas)`.
+3. The sender accumulates messages in typed collections.
+4. Every 1 second, `Flush()` sends all accumulated messages to each configured webhook endpoint as a batched JSON POST.
+5. Messages are filtered by area — a webhook configured for specific areas only receives messages from matching geofences.
+
+### Fort Change Webhooks
+
+Fort updates (new/edit/removal) go through `CreateFortWebHooks(old, new, change)`:
+- **NEW**: Sends new fort data. Triggered when `newRecord` is true.
+- **EDIT**: Compares old vs new for name, description, image URL (path-only comparison), and location (with float tolerance). Only sends if actual changes detected.
+- **REMOVAL**: Sends old fort data. Triggered by fort tracker stale detection.
+
+The `oldValues` for EDIT comparison come from `snapshotOldValues()` called at lock acquisition time.
+
+## In-Memory Fort Tracking
+
+### Purpose
+
+The fort tracker detects when forts are removed from the game or converted between types (pokestop ↔ gym), using S2 cell-level tracking of which forts exist.
+
+### Data Model
+
+```
+FortTracker
+  ├── cells: map[uint64]*FortTrackerCellState    // S2 cell → {lastSeen, pokestops set, gyms set}
+  └── forts: map[string]*FortTrackerLastSeen      // fort ID → {cellId, lastSeen, isGym}
+```
+
+### Detection Flow
+
+1. Each GMO response contains S2 cell IDs and the forts within them.
+2. `CheckRemovedForts()` calls `ProcessCellUpdate()` for each cell.
+3. For each cell, forts in the previous state but NOT in the current GMO are candidates for removal.
+4. If a fort has been missing for longer than `staleThreshold` (default 1 hour), it's marked as stale and deleted via `clearGymWithLock`/`clearPokestopWithLock`.
+5. If a pokestop ID appears in the gym set (or vice versa), it's detected as a type conversion — the old type is marked deleted (but not removed from the tracker).
+
+### API
+
+- `GET /api/fort-tracker/cell/:cell_id` — returns forts in an S2 cell
+- `GET /api/fort-tracker/forts/:fort_id` — returns a fort's cell and last-seen timestamp
+
+## Spatial Indexes (R-trees)
+
+### Two-Level Architecture
+
+Both pokemon and fort spatial queries use the same two-level pattern to avoid holding entity locks during search:
+
+1. **R-tree** (`rtree.RTreeG`): Maps `[lon, lat]` points to entity IDs. Provides fast bounding-box searches.
+2. **Lookup cache** (`xsync.MapOf`): Maps entity IDs to lightweight structs containing only the fields needed for filter matching. Lock-free concurrent reads.
+
+This separation means a scan of 100,000 pokemon in a bounding box only touches the R-tree and lookup cache — no entity mutexes are acquired until the final step of building API results for the matched subset.
+
+### Pokemon R-tree (`pokemonRtree.go`)
+
+**PokemonLookup** stores: PokemonId, Form, Weather, IVs (Atk/Def/Sta), Level, CP, Gender, Size, Iv percentage, HasEncounterValues flag. Uses `-1` sentinel for missing nullable values.
+
+**PokemonPvpLookup** stores: best PVP rank in Little/Great/Ultra leagues.
+
+**Lifecycle**:
+- Added to tree when pokemon is first saved or loaded from DB on cache miss (`pokemonRtreeUpdatePokemonOnGet`)
+- Updated on every save via `updatePokemonLookup()` which also recalculates PVP rankings
+- Removed via cache eviction callback when pokemon TTL expires
+
+### Fort R-tree (`fortRtree.go`)
+
+**FortLookup** stores a union of fields across all fort types: type indicator, location, power-up level, AR eligibility, plus type-specific fields:
+- **Gym**: team, slots, raid level/pokemon/timestamps
+- **Pokestop**: lure, quest rewards (both AR and non-AR), incident data, contest data
+- **Station**: battle level/pokemon/timestamps
+
+Enabled by `config.Config.FortInMemory`. Fort cache TTL is extended to 25 hours to keep entries resident.
+
+**Incident data** on FortLookup is updated separately via `updatePokestopIncidentLookup()` because incidents load after pokestops during preload, and incident updates come through a different code path than pokestop updates.
+
+### Scanning and DNF Filters
+
+#### Pokemon Scan
+
+Three API versions exist (V1/V2/V3), all following the same pattern:
+
+1. Copy the R-tree (read lock) for thread-safe traversal
+2. `rtree.Search(minLon, minLat, maxLon, maxLat)` to get candidate pokemon IDs
+3. For each ID, load `PokemonLookup` + `PokemonPvpLookup` from lookup cache
+4. Apply DNF filter matching
+5. Collect matching IDs up to a configurable limit
+6. For matched IDs, call `peekPokemonRecordReadOnly()` to lock and build full API results
+
+**DNF (Disjunctive Normal Form) Filters**: An array of filter clauses OR'd together. Each clause has AND'd conditions (IV range, level range, CP range, pokemon ID + form, PVP ranking, gender, size). A pokemon matches if ANY clause fully matches.
+
+**Filter lookup optimization**: Filters are pre-indexed by `{pokemonId, form}` key. For each pokemon, the system tries:
+1. Exact `{pokemonId, form}` match
+2. Wildcard form: `{pokemonId, -1}`
+3. Global catch-all: `{-1, -1}`
+
+This avoids iterating all filters for every pokemon.
+
+#### Fort Scan
+
+`internalGetForts()` and `internalGetFortsCombined()` follow the same pattern:
+
+1. Copy fort R-tree
+2. Bounding-box search for fort IDs
+3. Load `FortLookup` from lookup cache
+4. Apply `isFortDnfMatch()` which checks fort type, then type-specific fields:
+   - **Gym**: raid level, raid pokemon, raid expiry timestamp
+   - **Pokestop**: quest rewards (unified AR/non-AR matching), incidents, lures, contests
+   - **Station**: battle level, battle pokemon, battle expiry
+5. Lock and load full entity records for matched IDs
+
+The `FortCombinedScanEndpoint` scans all three fort types in one pass and splits results by type.
+
+## Preload
+
+On startup, `Preload()` bulk-loads entities from the database into caches. Order matters when `FortInMemory` is enabled:
+
+1. Pokestops → populates cache + fort R-tree
+2. Gyms → populates cache + fort R-tree
+3. Stations → populates cache + fort R-tree
+4. Incidents → **must load after pokestops** because `updatePokestopIncidentLookup` needs existing fort R-tree entries
+5. Pokemon → populates cache + pokemon R-tree
+
+Fort tracker is initialized from the preloaded fort data.