diff --git a/lib/streaming_providers/base/utils/mpd_cache.py b/lib/streaming_providers/base/utils/mpd_cache.py
index 09dee38..ba32ba8 100644
--- a/lib/streaming_providers/base/utils/mpd_cache.py
+++ b/lib/streaming_providers/base/utils/mpd_cache.py
@@ -76,8 +76,17 @@ class MPDCacheManager:
                 logger.debug(
                     f"Cache expired for {cache_key} (expired {staleness}s ago)"
                 )
-                self.vfs.delete(manifest_file)
-                self.vfs.delete(meta_file)
+                # Deliberately NOT deleting the expired files here. This get()
+                # call may have been made with the default max_stale=0 (a
+                # normal cache lookup), but a *different* caller further down
+                # the request chain may retry with max_stale>0 as a
+                # last-resort fallback after a live refetch fails. If we
+                # delete on every plain expiry check, that fallback caller
+                # always finds nothing — the fallback only ever worked when
+                # it happened to be the first caller to see the expired
+                # entry. Expired entries are reaped on their own schedule by
+                # clear_expired() instead, so both call patterns keep working
+                # regardless of ordering.
                 return None
 
             manifest_content = self.vfs.read_text(manifest_file)
@@ -264,4 +273,4 @@ class MPDCacheManager:
             return meta
         except Exception as e:
             logger.debug(f"Error getting cache info for {cache_key}: {e}")
-            return None
+            return None
\ No newline at end of file
diff --git a/lib/streaming_providers/base/utils/mpd_rewriter.py b/lib/streaming_providers/base/utils/mpd_rewriter.py
index 1871269..847b05b 100644
--- a/lib/streaming_providers/base/utils/mpd_rewriter.py
+++ b/lib/streaming_providers/base/utils/mpd_rewriter.py
@@ -7,6 +7,7 @@ Handles URL proxying, DRM key injection, quality filtering, and representation b
 import base64
 import struct
 import xml.etree.ElementTree as ET
+from dataclasses import dataclass
 from typing import Optional, Tuple, Set, Dict
 from urllib.parse import urljoin, quote, urlencode
 from datetime import datetime, timezone
@@ -18,6 +19,31 @@ from .drm_key_manager import KeyConfiguration
 from .representation_blocklist import RepresentationBlocklist
 from .video_quality import VideoQualityFilter, VideoRepresentation
 from .time_utils import parse_iso_duration
+
+
+@dataclass
+class _RewriteState:
+    """
+    Per-node traversal state for MPDRewriter._rewrite_node.
+
+    This used to be five separate positional parameters threaded through
+    every recursive call by hand. That's how the $RepresentationID$
+    substitution bug happened: current_rep_id was tracked as a local
+    variable in _rewrite_node but never added to the recursive call's
+    argument list, so it silently reset to None on every recursion step —
+    meaning it only ever "worked" for attributes living directly on the
+    <Representation> element itself, never for the <SegmentTemplate> child
+    where media/initialization templates actually live in practice.
+
+    Bundling the state into one object closes off that whole class of bug:
+    a recursive call either forwards the state object or it doesn't compile
+    (there's no way to forward "most of" a dataclass instance by accident).
+    """
+    base_url: str
+    period_id: str = ""
+    encrypted: bool = False
+    kid: Optional[str] = None
+    rep_id: Optional[str] = None
 from ..models.drm.constants import DRM_SYSTEM_NAMES
 
 
@@ -188,8 +214,9 @@ class MPDRewriter:
                 raise ValueError("No AdaptationSets remain after key filtering - manifest would be empty")
 
             # Rewrite URLs with appropriate keys and context-aware base URLs
-            self._rewrite_node(root, mpd_base_url, encrypted_ids, as_id_to_kid, base_url_map,
-                               False, None, "", best_video_info)
+            self._rewrite_node(
+                root, _RewriteState(base_url=mpd_base_url), encrypted_ids, as_id_to_kid, base_url_map
+            )
 
             rewritten = ET.tostring(root, encoding="unicode", method="xml")
             if not rewritten.startswith("<?xml"):
@@ -536,46 +563,52 @@ class MPDRewriter:
     def _rewrite_node(
             self,
             element: ET.Element,
-            base_url: str,
+            state: _RewriteState,
             encrypted_ids: Set[str],
             as_id_to_kid: Dict[str, str],
             base_url_map: Dict[str, str],
-            current_encrypted: bool,
-            current_kid: Optional[str] = None,
-            current_period_id: str = "",
-            best_video_info: Optional[VideoRepresentation] = None,
     ):
         """Recursive node rewriter with KID-aware key selection and context-aware base URLs."""
         # Track period ID as we traverse
         if element.tag.endswith("Period"):
-            current_period_id = element.get("id", "")
+            period_id = element.get("id", "")
+            base_url = state.base_url
 
             # When entering a new period, check if we have a period-level base URL stored
-            period_key = f"period_{current_period_id}" if current_period_id else "period_root"
+            period_key = f"period_{period_id}" if period_id else "period_root"
             if period_key in base_url_map:
                 # Update base_url to the period-specific base URL
                 base_url = base_url_map[period_key]
-#                logger.debug(f"Period {current_period_id} using stored base URL: {base_url}")
+#                logger.debug(f"Period {period_id} using stored base URL: {base_url}")
+
+            child_state = _RewriteState(
+                base_url=base_url,
+                period_id=period_id,
+                encrypted=state.encrypted,
+                kid=state.kid,
+                rep_id=state.rep_id,
+            )
 
             # Process all children of this period
             for child in list(element):
-                self._rewrite_node(
-                    child, base_url, encrypted_ids, as_id_to_kid, base_url_map,
-                    current_encrypted, current_kid, current_period_id, best_video_info
-                )
+                self._rewrite_node(child, child_state, encrypted_ids, as_id_to_kid, base_url_map)
             return  # Don't process further - we've handled all children
 
         # Track representation ID for template substitution
-        current_rep_id = None
+        rep_id = state.rep_id
         if element.tag.endswith("Representation"):
-            current_rep_id = element.get("id", "")
+            rep_id = element.get("id", "")
+
+        base_url = state.base_url
+        encrypted = state.encrypted
+        kid = state.kid
 
         # Update state when entering an AdaptationSet
         if element.tag.endswith("AdaptationSet"):
             as_id = element.get("id", str(id(element)))
             # Use same unique ID logic as _prepare_tree_and_extract_kids
-            unique_id = f"{current_period_id}_{as_id}" if current_period_id else as_id
-            current_encrypted = unique_id in encrypted_ids
+            unique_id = f"{state.period_id}_{as_id}" if state.period_id else as_id
+            encrypted = unique_id in encrypted_ids
 
             # Update base_url to the AdaptationSet-specific base URL
             if unique_id in base_url_map:
@@ -583,22 +616,33 @@ class MPDRewriter:
 #                logger.debug(f"AdaptationSet {unique_id} using stored base URL: {base_url}")
 
             # Get specific KID for this AdaptationSet (multi-key mode only)
-            if current_encrypted and not self.key_config.single_key_mode:
-                current_kid = as_id_to_kid.get(unique_id)
+            if encrypted and not self.key_config.single_key_mode:
+                kid = as_id_to_kid.get(unique_id)
+
+        # This is the state that applies to *this* element and everything
+        # below it — including rep_id, which is the piece that previously
+        # got dropped on recursion (see _RewriteState docstring).
+        current_state = _RewriteState(
+            base_url=base_url,
+            period_id=state.period_id,
+            encrypted=encrypted,
+            kid=kid,
+            rep_id=rep_id,
+        )
 
         # Handle BaseURL elements - THESE MUST BE REWRITTEN TO PROXY URLS
         if element.tag.endswith("BaseURL") and element.text:
             raw_url = element.text.strip()
             if raw_url:
                 # Resolve the relative BaseURL against the current base
-                resolved_cdn_url = self._urljoin_preserve_query(base_url, raw_url)
+                resolved_cdn_url = self._urljoin_preserve_query(current_state.base_url, raw_url)
 
                 # Rewrite the BaseURL text to a proxy URL
                 # In receiver-side clearkey mode, this will use /api/proxy/
                 element.text = self.build_proxy_url(
                     resolved_cdn_url, None, None,
-                    current_encrypted, current_kid,
-                    representation_id=current_rep_id
+                    current_state.encrypted, current_state.kid,
+                    representation_id=current_state.rep_id
                 )
 
         # ------------------------------------------------------------------
@@ -618,39 +662,36 @@ class MPDRewriter:
                 if not val:
                     continue
 
-                resolved = self._urljoin_preserve_query(base_url, val)
+                resolved = self._urljoin_preserve_query(current_state.base_url, val)
                 if "$" in resolved:
                     # Use shared utility for splitting template URLs
                     path, pattern = URLResolver.split_template_url(resolved)
                     element.attrib[attr] = self.build_proxy_url(
-                        path, pattern, seg_type, current_encrypted, current_kid,
-                        representation_id=current_rep_id
+                        path, pattern, seg_type, current_state.encrypted, current_state.kid,
+                        representation_id=current_state.rep_id
                     )
                 else:
                     element.attrib[attr] = self.build_proxy_url(
-                        resolved, None, seg_type, current_encrypted, current_kid,
-                        representation_id=current_rep_id
+                        resolved, None, seg_type, current_state.encrypted, current_state.kid,
+                        representation_id=current_state.rep_id
                     )
 
         # Handle SegmentURL (always 'media' type)
         if element.tag.endswith("SegmentURL") and "media" in element.attrib:
-            resolved = self._urljoin_preserve_query(base_url, element.attrib["media"])
+            resolved = self._urljoin_preserve_query(current_state.base_url, element.attrib["media"])
             path, pattern = (
                 URLResolver.split_template_url(resolved)
                 if "$" in resolved
                 else (resolved, None)
             )
             element.attrib["media"] = self.build_proxy_url(
-                path, pattern, "media", current_encrypted, current_kid,
-                representation_id=current_rep_id
+                path, pattern, "media", current_state.encrypted, current_state.kid,
+                representation_id=current_state.rep_id
             )
 
-        # Recurse to children
+        # Recurse to children, forwarding the full state (including rep_id)
         for child in element:
-            self._rewrite_node(
-                child, base_url, encrypted_ids, as_id_to_kid, base_url_map,
-                current_encrypted, current_kid, current_period_id, best_video_info
-            )
+            self._rewrite_node(child, current_state, encrypted_ids, as_id_to_kid, base_url_map)
 
     @staticmethod
     def _urljoin_preserve_query(base: str, url: str) -> str:
diff --git a/lib/streaming_providers/base/utils/vfs.py b/lib/streaming_providers/base/utils/vfs.py
index 4df930a..1ee758c 100644
--- a/lib/streaming_providers/base/utils/vfs.py
+++ b/lib/streaming_providers/base/utils/vfs.py
@@ -6,6 +6,7 @@ Provides transparent file operations for both Kodi and regular Python environmen
 
 import json
 import os
+import threading
 from typing import Any, Dict, List, Optional, Tuple
 
 # Import centralized environment manager
@@ -65,7 +66,9 @@ class VFS:
 
                 logger.info(f"Base path from environment: {self._base_path}")
 
-            # Ensure base directory exists
+            # Ensure base directory exists. self._base_path is already assigned
+            # above, so the join_path() call inside mkdirs() will not recurse
+            # back into this property.
             self.mkdirs("")
 
         return self._base_path
@@ -346,7 +349,7 @@ class VFS:
             pattern: File pattern filter (basic glob patterns)
 
         Returns:
-            List of filenames
+            List of filenames (basenames only, files only — not subdirectories)
         """
         try:
             if not dirpath:
@@ -384,6 +387,26 @@ class VFS:
             logger.error(f"Error listing files in {dirpath}: {e}")
             return []
 
+    def listdir(self, dirpath: str = "") -> List[str]:
+        """
+        Alias for list_files(dirpath) with the default "*" pattern.
+
+        Several call sites (e.g. MPDCacheManager.clear_all /
+        clear_expired) call self.vfs.listdir() rather than
+        self.vfs.list_files() — this previously did not exist on VFS at
+        all and raised AttributeError on every call, which was silently
+        swallowed by those callers' broad except blocks. Kept as an
+        explicit method (not just a `listdir = list_files` assignment)
+        so it shows up in stack traces under its own name.
+
+        Args:
+            dirpath: Directory path to list (relative to base_path)
+
+        Returns:
+            List of filenames (basenames only, files only)
+        """
+        return self.list_files(dirpath)
+
     def get_size(self, filepath: str) -> Optional[int]:
         """
         Get file size in bytes
@@ -471,8 +494,14 @@ class VFS:
         return info
 
 
-# Cache for VFS instances with different configurations
+# Cache for VFS instances with different configurations.
+# Guarded by a lock: this backend runs multi-threaded (Bottle), and without
+# it two threads racing on the same not-yet-cached key could each construct
+# and briefly use their own VFS instance for the same directory. Functionally
+# harmless (both point at the same base_path), but it defeats the point of
+# caching and is cheap to close off properly.
 _vfs_cache: Dict[Tuple[Optional[str], str], "VFS"] = {}
+_vfs_cache_lock = threading.Lock()
 
 
 def get_vfs(config_dir: Optional[str] = None, addon_subdir: str = "") -> "VFS":
@@ -488,14 +517,21 @@ def get_vfs(config_dir: Optional[str] = None, addon_subdir: str = "") -> "VFS":
     Returns:
         VFS instance
     """
-    global _vfs_cache
-
     cache_key = (config_dir, addon_subdir)
 
-    if cache_key not in _vfs_cache:
-        _vfs_cache[cache_key] = VFS(config_dir, addon_subdir)
+    # Fast path without the lock for the common case (already cached).
+    vfs = _vfs_cache.get(cache_key)
+    if vfs is not None:
+        return vfs
 
-    return _vfs_cache[cache_key]
+    with _vfs_cache_lock:
+        # Re-check inside the lock in case another thread created it while
+        # we were waiting.
+        vfs = _vfs_cache.get(cache_key)
+        if vfs is None:
+            vfs = VFS(config_dir, addon_subdir)
+            _vfs_cache[cache_key] = vfs
+        return vfs
 
 
 # For backward compatibility with existing code