feat: initial Luma implementation

Full self-hosted photo manager with:
- Ingestion worker (chokidar + sharp + exifr, dHash perceptual hashing)
- SQLite schema with WAL mode (photos, tags, collections, duplicate_pairs)
- Express API with keyset pagination, tag/collection/search/date filters
- React + Vite UI: timeline, lightbox, detail panel, tag/collection editors
- Duplicate detection (Union-Find grouping) with review + dismiss UI
- Map view (Leaflet + OpenStreetMap, marker clustering)
- Docker multi-stage build + Compose with inline env config
- GitHub Actions: CI build check + ghcr.io image publish on main/tags

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

worker/hash.js

@@ -0,0 +1,31 @@
+import sharp from 'sharp';
+
+// dHash (difference hash): 9×8 → 64-bit comparison → 16-char hex string.
+// Two images with Hamming distance ≤ 10 are near-duplicates.
+export async function computeDHash(filepath) {
+  const { data } = await sharp(filepath)
+    .resize(9, 8, { fit: 'fill' })
+    .greyscale()
+    .raw()
+    .toBuffer({ resolveWithObject: true });
+
+  let hash = 0n;
+  for (let row = 0; row < 8; row++) {
+    for (let col = 0; col < 8; col++) {
+      const left  = data[row * 9 + col];
+      const right = data[row * 9 + col + 1];
+      hash = (hash << 1n) | (left < right ? 1n : 0n);
+    }
+  }
+  return hash.toString(16).padStart(16, '0');
+}
+
+export function hammingDistance(a, b) {
+  let xor = BigInt('0x' + a) ^ BigInt('0x' + b);
+  let dist = 0;
+  while (xor > 0n) {
+    dist += Number(xor & 1n);
+    xor >>= 1n;
+  }
+  return dist;
+}