fix(upload): prevent false 422s and EMFILE crash during bulk uploads

Four related issues made uploading 271+ activities unreliable:

1. merge_all/write_combined_feed were inside the extraction try/except —
   any merge race returned 422 even though the file was on disk, causing
   the mobile app to permanently mark the upload as failed.  Fixed by
   moving them to a separate best-effort try/except after the extraction
   block.  Switch to merge_one (single-activity symlink) instead of
   merge_all (full rebuild) so each upload is O(1) FS ops, not O(N).

2. The dev watcher fired merge_all for every activity .json write AND the
   upload endpoint also ran merge_all — O(N²) symlink operations during
   bulk uploads.  Watcher now skips activities/*.json changes (upload
   endpoint handles those directly).

3. Vite/Chokidar followed the public/data symlink and opened a handle per
   activity file; constant merge rebuilds exhausted file descriptors and
   crashed the Astro dev server.  Fixed with watch.ignored on public/data.

4. _write_year_shards and write_combined_feed used f.unlink() without
   missing_ok=True — concurrent callers racing the same file threw
   FileNotFoundError which propagated as a false extraction failure.

site/astro.config.mjs

@@ -39,6 +39,12 @@ export default defineConfig({
     // Proxy /api/* to bincio serve/edit so cookies work same-origin in dev.
     // In production nginx handles this — same pattern, no code change needed.
     server: {
+      watch: {
+        // public/data is a symlink to the live data dir; Chokidar follows it and
+        // opens a handle per file, which causes EMFILE during bulk uploads as
+        // activity count grows. Data files don't need HMR — exclude them.
+        ignored: ['**/public/data/**'],
+      },
       proxy: {
         // Both /api/upload and /api/upload/strava-zip return SSE streams in response
         // to POST requests. Vite's default proxy buffers the full body before forwarding,