feat(download): one grab per Soulseek peer, several peers at once

slskd was capped at one transfer per daemon, on the grounds that
Soulseek peers punish clients that ask for too much. That politeness is
per peer: two different users do not compete for anyone's upload slot.
So one slow peer serialised every other Soulseek download behind it.

The manager now takes a per-(provider, peer) lock before any slot, so a
grab waiting on a busy peer does not hold a provider slot another peer
could use, and the slskd default rises to 3, which now counts peers.
The help text says so.

Running grabs at once exposed the folder collision: slskd names a
download's directory after the remote leaf folder, so two peers'
"Greatest Hits" (or any two rips' "CD1") share one directory, and
collect finds files by name there. Grabs whose local folders overlap
now take a package-level lock per folder, in sorted order, keyed on the
full path because two clients can share one daemon.

Closes #272

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017HJiuc3ZZhxsPXz3ozTirT

backend/download/provider_slskd.go

@@ -10,6 +10,7 @@ import (
 	"path"
 	"path/filepath"
 	"regexp"
+	"slices"
 	"strings"
 	"time"
 
@@ -742,6 +743,12 @@ func (s *slskd) Grab(
 	// first poll came back — an old failure failing a transfer that has
 	// not started.  So what is already terminal is noted before enqueueing
 	// and ignored after.
+	release, err := lockSlskdFolders(ctx, s.localFolders(c))
+	if err != nil {
+		return Result{}, err
+	}
+	defer release()
+
 	stale := s.terminalTransferIDs(ctx, username)
 
 	wanted := make([]map[string]any, 0, len(c.Files))
@@ -767,6 +774,59 @@ func (s *slskd) Grab(
 	return s.collect(c, dst)
 }
 
+// slskdFolders serialises grabs that land in the same local folder.
+//
+// slskd names a download's directory after the remote *leaf* folder, so
+// two different albums both shared as "Greatest Hits" — or any two
+// multi-disc rips, whose leaves are "CD1" and "CD2" — are written into
+// one directory, and collect finds files by name there.  Run at once,
+// a file one peer never sent is filled by the other peer's file of the
+// same name.  One grab per peer made that impossible; several peers at
+// once makes it likely.  It is package-level and keyed on the full
+// path because two configured clients can share one daemon.
+var slskdFolders keyedLock[string]
+
+// localFolders returns the directories under downloadsPath a candidate's
+// files will be written to, sorted so every grab takes them in the same
+// order and two cannot each hold what the other waits for.
+func (s *slskd) localFolders(c Candidate) []string {
+	var out []string
+
+	for _, f := range c.Files {
+		norm := strings.ReplaceAll(f.Path, `\`, "/")
+		out = append(out, filepath.Join(s.downloadsPath, path.Base(path.Dir(norm))))
+	}
+
+	slices.Sort(out)
+
+	return slices.Compact(out)
+}
+
+// lockSlskdFolders takes every folder in order, releasing what it holds
+// if the context ends part way.
+func lockSlskdFolders(ctx context.Context, folders []string) (func(), error) {
+	releases := make([]func(), 0, len(folders))
+
+	releaseAll := func() {
+		for _, r := range slices.Backward(releases) {
+			r()
+		}
+	}
+
+	for _, f := range folders {
+		r, err := slskdFolders.acquire(ctx, f)
+		if err != nil {
+			releaseAll()
+
+			return nil, err
+		}
+
+		releases = append(releases, r)
+	}
+
+	return releaseAll, nil
+}
+
 // slskdDownloadsPath is the transfers endpoint for one peer.  Soulseek
 // usernames may contain spaces and punctuation, so the name is escaped
 // rather than spliced into the path.