diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index 229a4d599..5d70ee765 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -256,3 +256,10 @@ jobs:
         env:
           PYTEST_SHARD: ${{ matrix.shard }}
         run: python -m pytest -q -rs --durations=25 --shard "$PYTEST_SHARD"
+
+      - name: Playwright mobile layout checks (shard 1/4 only)
+        if: steps.docs-check.outputs.docs_only != 'true' && matrix.shard == '1/4'
+        run: |
+          ODYSSEUS_TEST_PYTHON=python ODYSSEUS_LAYOUT_BROWSER=chromium npx --no-install playwright test \
+            --config tests/e2e/ios-layout.config.cjs \
+            --reporter=line
diff --git a/package.json b/package.json
index aa454e017..79cf83c3d 100644
--- a/package.json
+++ b/package.json
@@ -9,7 +9,11 @@
     "test:photo-editor": "playwright test --config tests/e2e/playwright.config.js",
     "test:photo-editor:install": "playwright install chromium firefox webkit",
     "test:photo-editor:firefox": "PHOTO_EDITOR_E2E_BROWSER=firefox playwright test --config tests/e2e/playwright.config.js",
-    "test:photo-editor:webkit": "PHOTO_EDITOR_E2E_BROWSER=webkit playwright test --config tests/e2e/playwright.config.js"
+    "test:photo-editor:webkit": "PHOTO_EDITOR_E2E_BROWSER=webkit playwright test --config tests/e2e/playwright.config.js",
+    "test:ios-layout": "playwright test --config tests/e2e/ios-layout.config.cjs",
+    "test:ios-layout:chromium": "ODYSSEUS_LAYOUT_BROWSER=chromium playwright test --config tests/e2e/ios-layout.config.cjs",
+    "test:ios-layout:webkit": "ODYSSEUS_LAYOUT_BROWSER=webkit playwright test --config tests/e2e/ios-layout.config.cjs",
+    "test:ios-layout:install": "playwright install chromium webkit"
   },
   "devDependencies": {
     "@antithesishq/bombadil": "^0.7.0",
diff --git a/routes/cookbook_helpers.py b/routes/cookbook_helpers.py
index 856c8bdb5..24419c29c 100644
--- a/routes/cookbook_helpers.py
+++ b/routes/cookbook_helpers.py
@@ -1030,6 +1030,166 @@ def _append_llama_cpp_linux_accel_build_lines(runner_lines: list[str]) -> None:
     runner_lines.append('  fi  # end _odysseus_have_prebuilt guard')
 
 
+ODYSSEUS_LLAMA_SHIM_MARKER = "# Auto-generated by Odysseus Cookbook: a `llama-server` lookalike"
+
+
+def _append_llama_server_resolution_definition_lines(runner_lines: list[str]) -> None:
+    """Define the helper shell function that discovers, reconciles and prioritizes llama-server."""
+    marker = f"'{_bash_squote(ODYSSEUS_LLAMA_SHIM_MARKER)}'"
+    runner_lines.append('_odysseus_resolve_llama_server() {')
+    runner_lines.append('  _odysseus_llama_bin=""')
+    runner_lines.append(f'  local _marker={marker}')
+    runner_lines.append('  local _saved_ifs="$IFS"')
+    runner_lines.append("  IFS=':'")
+    runner_lines.append('  local -a _parts')
+    runner_lines.append('  read -r -a _parts <<< "$PATH"')
+    runner_lines.append('  IFS="$_saved_ifs"')
+    runner_lines.append('  local _dir _cand')
+    runner_lines.append('  for _dir in "${_parts[@]}"; do')
+    runner_lines.append('    [ -n "$_dir" ] || continue')
+    runner_lines.append('    _cand="$_dir/llama-server"')
+    runner_lines.append('    [ -x "$_cand" ] || continue')
+    runner_lines.append('    if [ -f "$_cand" ] && [ ! -L "$_cand" ] && grep -Fxq "$_marker" "$_cand" 2>/dev/null; then')
+    runner_lines.append('      continue')
+    runner_lines.append('    fi')
+    runner_lines.append('    if "$_cand" --version >/dev/null 2>&1; then')
+    runner_lines.append('      _odysseus_llama_bin="$_cand"')
+    runner_lines.append('      break')
+    runner_lines.append('    fi')
+    runner_lines.append('  done')
+    runner_lines.append('  if [ -f "$HOME/bin/llama-server" ] && [ ! -L "$HOME/bin/llama-server" ] && grep -Fxq "$_marker" "$HOME/bin/llama-server" 2>/dev/null; then')
+    runner_lines.append('    if [ -n "$_odysseus_llama_bin" ]; then')
+    runner_lines.append('      echo "[odysseus] Native llama-server detected — retiring Odysseus compatibility shim in ~/bin"')
+    runner_lines.append('      rm -f "$HOME/bin/llama-server"')
+    runner_lines.append('      hash -r')
+    runner_lines.append('    elif python3 -c "import llama_cpp.server" 2>/dev/null; then')
+    runner_lines.append('      _odysseus_llama_bin="$HOME/bin/llama-server"')
+    runner_lines.append('    else')
+    runner_lines.append('      echo "[odysseus] Removing stale Odysseus llama-server shim (python3 -m llama_cpp.server unavailable)"')
+    runner_lines.append('      rm -f "$HOME/bin/llama-server"')
+    runner_lines.append('      hash -r')
+    runner_lines.append('    fi')
+    runner_lines.append('  fi')
+    runner_lines.append('  if [ -n "$_odysseus_llama_bin" ]; then')
+    runner_lines.append('    local _cand_dir')
+    runner_lines.append('    _cand_dir="$(dirname "$_odysseus_llama_bin")"')
+    runner_lines.append('    if [ "${PATH%%:*}" != "$_cand_dir" ]; then')
+    runner_lines.append('      export PATH="$_cand_dir:$PATH"')
+    runner_lines.append('    fi')
+    runner_lines.append('    hash -r')
+    runner_lines.append('    return 0')
+    runner_lines.append('  fi')
+    runner_lines.append('  return 1')
+    runner_lines.append('}')
+
+
+def _append_llama_server_shim_definition_lines(runner_lines: list[str]) -> None:
+    """Define the helper shell function that generates the Odysseus compatibility shim."""
+    _append_llama_server_resolution_definition_lines(runner_lines)
+    marker = f"'{_bash_squote(ODYSSEUS_LLAMA_SHIM_MARKER)}'"
+    runner_lines.append('_odysseus_create_llama_shim() {')
+    runner_lines.append('  local _target="$HOME/bin/llama-server"')
+    runner_lines.append(f'  local _marker={marker}')
+    runner_lines.append('  if [ -e "$_target" ] || [ -L "$_target" ]; then')
+    runner_lines.append('    if [ -f "$_target" ] && [ ! -L "$_target" ] && grep -Fxq "$_marker" "$_target" 2>/dev/null; then')
+    runner_lines.append('      : # existing Odysseus shim — safe to replace')
+    runner_lines.append('    else')
+    runner_lines.append('      echo "[odysseus] WARNING: $_target already exists and is not an Odysseus-managed shim; leaving untouched." >&2')
+    runner_lines.append('      return 1')
+    runner_lines.append('    fi')
+    runner_lines.append('  fi')
+    runner_lines.append('  mkdir -p "$HOME/bin" || return 1')
+    runner_lines.append('  local _tmp')
+    runner_lines.append('  _tmp="$(mktemp "$HOME/bin/.llama-server.odysseus.tmp.XXXXXX")" || {')
+    runner_lines.append('    echo "[odysseus] ERROR: unable to create secure temporary llama-server shim in $HOME/bin" >&2')
+    runner_lines.append('    return 1')
+    runner_lines.append('  }')
+    runner_lines.append("  if ! cat > \"$_tmp\" <<'_ODY_LLAMA_SHIM_EOF'")
+    runner_lines.append('#!/usr/bin/env bash')
+    runner_lines.append(ODYSSEUS_LLAMA_SHIM_MARKER)
+    runner_lines.append('# that translates the native CLI to `python -m llama_cpp.server`.')
+    runner_lines.append('# Lets cookbook-generated launch commands run unchanged on hosts')
+    runner_lines.append('# where only the pip llama-cpp-python package is installed.')
+    runner_lines.append('ARGS=()')
+    runner_lines.append('while [ $# -gt 0 ]; do')
+    runner_lines.append('  case "$1" in')
+    runner_lines.append('    -ngl|--gpu-layers|--n-gpu-layers) ARGS+=(--n_gpu_layers "$2"); shift 2 ;;')
+    runner_lines.append('    -c|--ctx-size) ARGS+=(--n_ctx "$2"); shift 2 ;;')
+    runner_lines.append('    -b|--batch-size) ARGS+=(--n_batch "$2"); shift 2 ;;')
+    runner_lines.append('    -ub|--ubatch-size) shift 2 ;;  # llama-cpp-python has no separate ubatch')
+    runner_lines.append('    --flash-attn) ARGS+=(--flash_attn true); shift 2 ;;')
+    runner_lines.append('    --cache-type-k) ARGS+=(--type_k "$2"); shift 2 ;;')
+    runner_lines.append('    --cache-type-v) ARGS+=(--type_v "$2"); shift 2 ;;')
+    runner_lines.append('    --n-cpu-moe) ARGS+=(--n_cpu_moe "$2"); shift 2 ;;')
+    runner_lines.append('    --mmproj) ARGS+=(--clip_model_path "$2"); shift 2 ;;')
+    runner_lines.append('    --image-max-tokens) shift 2 ;;  # native-only')
+    runner_lines.append('    --no-mmap) ARGS+=(--no_mmap true); shift ;;')
+    runner_lines.append('    --no-warmup) shift ;;  # native-only')
+    runner_lines.append('    --chat-template) ARGS+=(--chat_format "$2"); shift 2 ;;')
+    runner_lines.append('    --fit|--split-mode|--tensor-split|--main-gpu|--parallel) shift 2 ;;  # native-only')
+    runner_lines.append('    --mlock) ARGS+=(--use_mlock true); shift ;;')
+    runner_lines.append('    *) ARGS+=("$1"); shift ;;')
+    runner_lines.append('  esac')
+    runner_lines.append('done')
+    runner_lines.append('exec python3 -m llama_cpp.server "${ARGS[@]}"')
+    runner_lines.append('_ODY_LLAMA_SHIM_EOF')
+    runner_lines.append('  then')
+    runner_lines.append('    rm -f -- "$_tmp"')
+    runner_lines.append('    return 1')
+    runner_lines.append('  fi')
+    runner_lines.append('  chmod +x "$_tmp" || { rm -f -- "$_tmp"; return 1; }')
+    runner_lines.append('  mv -f -- "$_tmp" "$_target" || { rm -f -- "$_tmp"; return 1; }')
+    runner_lines.append('  hash -r')
+    runner_lines.append('  echo "[odysseus] Created llama-server shim → python -m llama_cpp.server (no native binary needed)"')
+    runner_lines.append('}')
+
+
+def _append_llama_server_shim_reconciliation_lines(runner_lines: list[str]) -> None:
+    """Reconcile an existing ~/bin/llama-server if it is owned by Odysseus Cookbook.
+
+    Invariants:
+    - Never touch an arbitrary user-owned executable or symlink.
+    - If ~/bin/llama-server is an Odysseus generated shim:
+        - If a native llama-server is present and healthy elsewhere on PATH, remove the shim
+          so the native binary is used.
+        - If python3 -m llama_cpp.server is unavailable, remove the stale shim
+          and hash -r so the fallback/build path is triggered.
+        - If no healthy native server exists and llama_cpp.server is working, preserve it.
+    - If a healthy native candidate exists on PATH (even later in PATH), prioritize it and
+      ensure the launched executable is the validated candidate.
+    """
+    runner_lines.append('[ -r ~/.config/odysseus-llama-cpp-env ] && . ~/.config/odysseus-llama-cpp-env')
+    runner_lines.append('if ! declare -F _odysseus_resolve_llama_server >/dev/null 2>&1; then')
+    _append_llama_server_resolution_definition_lines(runner_lines)
+    runner_lines.append('fi')
+    runner_lines.append('_odysseus_resolve_llama_server || true')
+
+
+def _append_llama_server_finalization_lines(runner_lines: list[str], platform_hint: str = "") -> None:
+    """Reconcile post-install launcher availability and enforce fail-closed invariant.
+
+    Guarantees that if python3 -m llama_cpp.server is usable but the `llama-server`
+    command is unresolvable, the compatibility launcher is created. If no usable
+    launcher exists after all build/install attempts, sets ODYSSEUS_PREFLIGHT_EXIT=127.
+    """
+    err_msg = (
+        f"ERROR: llama.cpp serving is not available on {platform_hint} after install attempts."
+        if platform_hint
+        else "ERROR: llama.cpp serving is not available after install/build attempts."
+    )
+    runner_lines.append('if ! declare -F _odysseus_resolve_llama_server >/dev/null 2>&1; then')
+    _append_llama_server_resolution_definition_lines(runner_lines)
+    runner_lines.append('fi')
+    runner_lines.append('if [ -z "$_odysseus_llama_bin" ] && python3 -c "import llama_cpp.server" 2>/dev/null; then')
+    runner_lines.append('  _odysseus_create_llama_shim')
+    runner_lines.append('  _odysseus_resolve_llama_server || true')
+    runner_lines.append('fi')
+    runner_lines.append('if [ -z "$_odysseus_llama_bin" ]; then')
+    runner_lines.append(f'  echo "{err_msg}"')
+    runner_lines.append('  ODYSSEUS_PREFLIGHT_EXIT=127')
+    runner_lines.append('fi')
+
+
 def _llama_cpp_rebuild_cmd(update_source: bool = False) -> str:
     """Shell command that clears the Cookbook-managed llama.cpp build.
 
diff --git a/routes/cookbook_routes.py b/routes/cookbook_routes.py
index 02b419894..ae2667f65 100644
--- a/routes/cookbook_routes.py
+++ b/routes/cookbook_routes.py
@@ -57,9 +57,11 @@ from routes.cookbook_helpers import (
     _pip_install_no_cache, _user_shell_path_bootstrap, _venv_safe_local_pip_install_cmd,
     _diagnose_serve_output, run_ssh_command_async,
     _ollama_bind_from_cmd, _pip_install_fallback_chain, _pip_install_no_cache,
-    _user_shell_path_bootstrap, _venv_safe_local_pip_install_cmd,
     _append_pip_install_runner_lines, _pip_install_command_without_break_system_packages,
     _normalize_llama_cpp_python_cache_types,
+    _append_llama_server_shim_definition_lines,
+    _append_llama_server_shim_reconciliation_lines,
+    _append_llama_server_finalization_lines,
     ModelDownloadRequest, ServeRequest,
 )
 
@@ -2161,7 +2163,7 @@ def setup_cookbook_routes() -> APIRouter:
                 ps_lines.append('}')
             elif "llama_cpp" in req.cmd or "llama-server" in req.cmd:
                 ps_lines.append('# Auto-install llama-cpp-python if missing')
-                ps_lines.append('try { python -c "import llama_cpp" 2>$null } catch {}')
+                ps_lines.append('try { python -c "import llama_cpp.server" 2>$null } catch {}')
                 ps_lines.append('if ($LASTEXITCODE -ne 0) {')
                 ps_lines.append('  Write-Host "Installing llama-cpp-python..."')
                 ps_lines.append('  python -m pip install llama-cpp-python[server]')
@@ -2247,14 +2249,17 @@ def setup_cookbook_routes() -> APIRouter:
                 # ollama is found (otherwise macOS falls back to a slow source build).
                 # /opt/homebrew = Apple Silicon, /usr/local = Intel; harmless on Linux.
                 runner_lines.append('export PATH="$HOME/.local/bin:$HOME/bin:$HOME/llama.cpp/build/bin:/opt/homebrew/bin:/usr/local/bin:$PATH"')
+                _append_llama_server_shim_definition_lines(runner_lines)
+                _append_llama_server_shim_reconciliation_lines(runner_lines)
                 runner_lines.append('if [ -d /data/data/com.termux ]; then')
                 runner_lines.append('  # Termux: no native build — use the Python bindings (CPU).')
-                runner_lines.append('  if ! python3 -c "import llama_cpp" 2>/dev/null; then')
+                runner_lines.append('  if ! python3 -c "import llama_cpp.server" 2>/dev/null; then')
                 runner_lines.append('    pkg install -y cmake 2>/dev/null')
                 runner_lines.append('    pip install numpy diskcache jinja2 2>/dev/null')
                 runner_lines.append('    CMAKE_ARGS="-DGGML_BLAS=OFF -DGGML_LLAMAFILE=OFF" pip install \'llama-cpp-python[server]\' --no-build-isolation --no-cache-dir 2>&1 || true')
                 runner_lines.append('  fi')
-                runner_lines.append('elif ! command -v llama-server &>/dev/null; then')
+                _append_llama_server_finalization_lines(runner_lines, platform_hint="Termux")
+                runner_lines.append('elif [ -z "$_odysseus_llama_bin" ]; then')
                 runner_lines.append('  echo "Native llama-server not found — building from source (one-time, may take a few minutes)..."')
                 runner_lines.append('  mkdir -p ~/bin')
                 runner_lines.append('  cd ~ && [ -d llama.cpp ] || git clone --depth 1 https://github.com/ggml-org/llama.cpp')
@@ -2271,8 +2276,7 @@ def setup_cookbook_routes() -> APIRouter:
                 # would reuse the bad settings and fail again. CMAKE_BUILD_TYPE is
                 # explicit so the binary is optimized (Metal auto-enables on macOS).
                 runner_lines.append('    cd ~/llama.cpp && rm -rf build && cmake -B build -DCMAKE_BUILD_TYPE=Release \\')
-                runner_lines.append('      && cmake --build build -j"$NPROC" --target llama-server \\')
-                runner_lines.append('      && ln -sf ~/llama.cpp/build/bin/llama-server ~/bin/llama-server')
+                runner_lines.append('      && cmake --build build -j"$NPROC" --target llama-server')
                 runner_lines.append('  else')
                 _append_llama_cpp_linux_accel_build_lines(runner_lines)
                 runner_lines.append('  fi')
@@ -2296,58 +2300,13 @@ def setup_cookbook_routes() -> APIRouter:
                 runner_lines.append('      fi')
                 runner_lines.append('    fi')
                 runner_lines.append('  fi')
-                # SHORT-CIRCUIT before the build/pip fallback: if the
-                # native binary is missing but llama_cpp Python is already
-                # installed, drop a wrapper at ~/bin/llama-server that
-                # translates llama-server CLI args to llama_cpp.server's
-                # underscore-style flags. The user's serve command stays
-                # `llama-server ...` and "just works" — no build, no cmake,
-                # no second install. This is the path that unblocks every
-                # remote where pip-installed llama-cpp-python is already
-                # working but Cookbook used to insist on a native binary.
-                runner_lines.append('  if ! command -v llama-server >/dev/null 2>&1 && python3 -c "import llama_cpp" 2>/dev/null; then')
-                runner_lines.append('    mkdir -p ~/bin')
-                runner_lines.append('    cat > ~/bin/llama-server <<\'_ODY_LLAMA_SHIM_EOF\'')
-                runner_lines.append('#!/usr/bin/env bash')
-                runner_lines.append('# Auto-generated by Odysseus Cookbook: a `llama-server` lookalike')
-                runner_lines.append('# that translates the native CLI to `python -m llama_cpp.server`.')
-                runner_lines.append('# Lets cookbook-generated launch commands run unchanged on hosts')
-                runner_lines.append('# where only the pip llama-cpp-python package is installed.')
-                runner_lines.append('ARGS=()')
-                runner_lines.append('while [ $# -gt 0 ]; do')
-                runner_lines.append('  case "$1" in')
-                runner_lines.append('    -ngl|--gpu-layers|--n-gpu-layers) ARGS+=(--n_gpu_layers "$2"); shift 2 ;;')
-                runner_lines.append('    -c|--ctx-size) ARGS+=(--n_ctx "$2"); shift 2 ;;')
-                runner_lines.append('    -b|--batch-size) ARGS+=(--n_batch "$2"); shift 2 ;;')
-                runner_lines.append('    -ub|--ubatch-size) shift 2 ;;  # llama-cpp-python has no separate ubatch')
-                runner_lines.append('    --flash-attn) ARGS+=(--flash_attn true); shift 2 ;;')
-                runner_lines.append('    --cache-type-k) ARGS+=(--type_k "$2"); shift 2 ;;')
-                runner_lines.append('    --cache-type-v) ARGS+=(--type_v "$2"); shift 2 ;;')
-                runner_lines.append('    --n-cpu-moe) ARGS+=(--n_cpu_moe "$2"); shift 2 ;;')
-                runner_lines.append('    --mmproj) ARGS+=(--clip_model_path "$2"); shift 2 ;;')
-                runner_lines.append('    --image-max-tokens) shift 2 ;;  # native-only')
-                runner_lines.append('    --no-mmap) ARGS+=(--no_mmap true); shift ;;')
-                runner_lines.append('    --no-warmup) shift ;;  # native-only')
-                runner_lines.append('    --chat-template) ARGS+=(--chat_format "$2"); shift 2 ;;')
-                runner_lines.append('    --fit|--split-mode|--tensor-split|--main-gpu|--parallel) shift 2 ;;  # native-only')
-                runner_lines.append('    --mlock) ARGS+=(--use_mlock true); shift ;;')
-                runner_lines.append('    *) ARGS+=("$1"); shift ;;')
-                runner_lines.append('  esac')
-                runner_lines.append('done')
-                runner_lines.append('exec python3 -m llama_cpp.server "${ARGS[@]}"')
-                runner_lines.append('_ODY_LLAMA_SHIM_EOF')
-                runner_lines.append('    chmod +x ~/bin/llama-server')
-                runner_lines.append('    echo "[odysseus] Created llama-server shim → python -m llama_cpp.server (no native binary needed)"')
-                runner_lines.append('  fi')
-                runner_lines.append('  # If the native build failed, fall back to the Python bindings.')
-                runner_lines.append('  if ! command -v llama-server &>/dev/null && ! python3 -c "import llama_cpp" 2>/dev/null; then')
+                runner_lines.append('  _odysseus_resolve_llama_server || true')
+                # If the native build failed, fall back to the Python bindings.
+                runner_lines.append('  if [ -z "$_odysseus_llama_bin" ] && ! python3 -c "import llama_cpp.server" 2>/dev/null; then')
                 runner_lines.append('    echo "llama-server build failed — installing Python bindings as fallback..."')
                 runner_lines.append(f"    {_pip_install_fallback_chain('llama-cpp-python[server]', python_cmd='pip')} || true")
                 runner_lines.append('  fi')
-                runner_lines.append('  if ! command -v llama-server &>/dev/null && ! python3 -c "import llama_cpp" 2>/dev/null; then')
-                runner_lines.append('    echo "ERROR: llama.cpp serving is not available after install/build attempts."')
-                runner_lines.append('    ODYSSEUS_PREFLIGHT_EXIT=127')
-                runner_lines.append('  fi')
+                _append_llama_server_finalization_lines(runner_lines)
                 runner_lines.append('fi')
             elif re.search(r"\bollama\s+serve\b", req.cmd):
                 handled_ollama_serve = True
diff --git a/routes/session_routes.py b/routes/session_routes.py
index 1c903f194..388de5b20 100644
--- a/routes/session_routes.py
+++ b/routes/session_routes.py
@@ -1315,7 +1315,7 @@ def setup_session_routes(
     def auto_sort_sessions(request: Request, skip_llm: bool = False):
         """Use AI to categorize all sessions into folders.
 
-        Phase 1 deletes empty/throwaway sessions and Phase 2 asks the LLM
+        Phase 1 deletes empty/Incognito sessions and Phase 2 asks the LLM
         to assign folders. When `skip_llm=true` the endpoint returns
         after Phase 1 — used by the "Tidy (no AI)" UI affordance so
         users can clean junk without spending tokens.
@@ -1326,21 +1326,11 @@ def setup_session_routes(
         single_user_mode = not user and _auth_disabled()
         user_sessions = session_manager.get_sessions_for_user(user)
 
-        # Delete empty and throwaway sessions before sorting
+        # Saved conversations are never disposable based on their title or prompt.
         from core.database import ChatMessage as DbMsg
         db = SessionLocal()
         deleted_empty = 0
         deleted_throwaway = 0
-        # Names that indicate a throwaway/test session (case-insensitive exact or prefix match)
-        _THROWAWAY_NAMES = {
-            "test", "testing", "asdf", "asd", "hello", "hi", "hey",
-            "yo", "sup", "hola", "hii", "hiii", "heyo",
-            "foo", "bar", "baz", "tmp", "temp", "scratch", "untitled",
-            "new chat", "delete", "remove", "junk", "trash", "xxx",
-            "abc", "qwerty", "blah", "stuff", "whatever", "idk",
-            "ok", "lol", "bruh", "hmm", "hm", "meh",
-        }
-        _THROWAWAY_MAX_MESSAGES = 4  # only delete if <= this many messages
         try:
             rows_q = db.query(DbSession).filter(DbSession.archived == False)
             if user:
@@ -1349,15 +1339,11 @@ def setup_session_routes(
                 rows_q = rows_q.filter(DbSession.owner == user)
             rows = rows_q.limit(2000).all()
             folder_map = {r.id: r.folder for r in rows}
-            # Precompute per-session message counts in TWO aggregate queries
+            # Precompute per-session message counts in one aggregate query
             # instead of 1–3 queries PER session — with many chats the per-row
             # loop was doing thousands of round-trips and blowing the timeout.
             from sqlalchemy import func as _sa_func
             _counts = dict(db.query(DbMsg.session_id, _sa_func.count(DbMsg.id)).group_by(DbMsg.session_id).all())
-            _asst_counts = dict(
-                db.query(DbMsg.session_id, _sa_func.count(DbMsg.id))
-                .filter(DbMsg.role == "assistant").group_by(DbMsg.session_id).all()
-            )
             cleanup_now = utcnow_naive()
             for row in rows:
                 # Never delete important sessions
@@ -1367,9 +1353,10 @@ def setup_session_routes(
                 if (row.name or "").strip() == "Incognito":
                     should_delete = True
                     deleted_throwaway += 1
-                    db.delete(row)
                     if hasattr(session_manager, 'delete_session'):
                         session_manager.delete_session(row.id)
+                    else:
+                        db.delete(row)
                     continue
                 if is_session_recently_active(row, now=cleanup_now):
                     continue
@@ -1378,31 +1365,11 @@ def setup_session_routes(
                 if msg_count == 0:
                     should_delete = True
                     deleted_empty += 1
-                elif msg_count <= _THROWAWAY_MAX_MESSAGES:
-                    name = (row.name or "").strip().lower()
-                    # Check first user message content (AI renames sessions, so
-                    # "hi" becomes "Casual Greeting Exchange" — name alone won't match)
-                    first_msg = db.query(DbMsg.content).filter(
-                        DbMsg.session_id == row.id, DbMsg.role == "user"
-                    ).order_by(DbMsg.timestamp).first()
-                    first_text = (first_msg[0] or "").strip().lower() if first_msg else ""
-                    # Count assistant messages — if user sent something but AI never replied, it's dead
-                    assistant_count = _asst_counts.get(row.id, 0)
-                    if name in _THROWAWAY_NAMES or name.startswith("chat:") or first_text in _THROWAWAY_NAMES:
-                        should_delete = True
-                        deleted_throwaway += 1
-                    # Single user message with no AI response = dead session
-                    elif msg_count == 1 and assistant_count == 0:
-                        should_delete = True
-                        deleted_throwaway += 1
-                    # Short phrase (1-3 words) with no real AI conversation (<=2 msgs)
-                    elif msg_count <= 2 and first_text and len(first_text.split()) <= 3 and len(first_text) <= 40:
-                        should_delete = True
-                        deleted_throwaway += 1
                 if should_delete:
-                    db.delete(row)
                     if hasattr(session_manager, 'delete_session'):
                         session_manager.delete_session(row.id)
+                    else:
+                        db.delete(row)
             if deleted_empty or deleted_throwaway:
                 db.commit()
                 logger.info(f"Auto-sort: deleted {deleted_empty} empty + {deleted_throwaway} throwaway sessions")
diff --git a/src/mcp_manager.py b/src/mcp_manager.py
index 53112009d..b47d0e1d7 100644
--- a/src/mcp_manager.py
+++ b/src/mcp_manager.py
@@ -10,6 +10,7 @@ import logging
 import os
 import re
 import asyncio 
+import contextlib
 from typing import Any, Dict, List, Optional, Set, Tuple
 from src.database import McpServer, SessionLocal
 
@@ -383,12 +384,21 @@ class McpManager:
 
             provider = build_provider(server_id, url, on_redirect=_on_redirect)
             stack = AsyncExitStack()
-            transport = await stack.enter_async_context(streamablehttp_client(url, auth=provider))
-            read_stream, write_stream, _get_session_id = transport
-            session = await stack.enter_async_context(ClientSession(read_stream, write_stream))
-            await session.initialize()
+            try:
+                transport = await stack.enter_async_context(streamablehttp_client(url, auth=provider))
+                read_stream, write_stream, _get_session_id = transport
+                session = await stack.enter_async_context(ClientSession(read_stream, write_stream))
+                await session.initialize()
 
-            tools_result = await session.list_tools()
+                tools_result = await session.list_tools()
+            except BaseException:
+                # Close the half-open transport here, in the task that entered
+                # it. Left for the garbage collector, its AnyIO cancel scope is
+                # exited from the wrong task and the event loop keeps spinning
+                # (#5518). Its own teardown error would only mask the real one.
+                with contextlib.suppress(Exception):
+                    await stack.aclose()
+                raise
             tools = []
             for tool in tools_result.tools:
                 tools.append({
diff --git a/src/session_actions.py b/src/session_actions.py
index 1efa484e3..9859ef486 100644
--- a/src/session_actions.py
+++ b/src/session_actions.py
@@ -12,16 +12,6 @@ from datetime import datetime, timedelta, timezone
 
 logger = logging.getLogger(__name__)
 
-# Names that indicate a throwaway/test session
-_THROWAWAY_NAMES = {
-    "test", "testing", "asdf", "asd", "hello", "hi", "hey",
-    "yo", "sup", "hola", "hii", "hiii", "heyo",
-    "foo", "bar", "baz", "tmp", "temp", "scratch", "untitled",
-    "new chat", "delete", "remove", "junk", "trash", "xxx",
-    "abc", "qwerty", "blah", "stuff", "whatever", "idk",
-    "ok", "lol", "bruh", "hmm", "hm", "meh",
-}
-_THROWAWAY_MAX_MESSAGES = 4
 _FRESH_EMPTY_SESSION_GRACE = timedelta(minutes=10)
 _FRESH_SESSION_GRACE = _FRESH_EMPTY_SESSION_GRACE
 
@@ -54,14 +44,15 @@ def is_session_recently_active(row, now=None, grace=_FRESH_SESSION_GRACE) -> boo
 
 
 async def run_auto_sort(owner: str, skip_llm: bool = False, delete_throwaway: bool = True) -> str:
-    """Run session cleanup + (optional) AI folder sort for the given owner.
+    """Remove empty/Incognito sessions and optionally organize saved chats.
 
     Args:
         owner: user whose sessions to process
-        skip_llm: when True, do only Phase 1 (delete empty/throwaway sessions);
+        skip_llm: when True, do only Phase 1 (delete empty/Incognito sessions);
             skip Phase 2 (AI folder assignment). Used by the built-in daily
             background sweep so it never burns LLM tokens.
-        delete_throwaway: when False, only empty/incognito sessions are deleted.
+        delete_throwaway: retained for caller compatibility. Ordinary sessions
+            with saved messages are always preserved, regardless of this flag.
 
     Returns a human-readable summary of what was done.
     """
@@ -71,7 +62,7 @@ async def run_auto_sort(owner: str, skip_llm: bool = False, delete_throwaway: bo
 
     db = SessionLocal()
     try:
-        # ── Phase 1: Delete empty/throwaway sessions ──
+        # ── Phase 1: Delete only empty and explicitly Incognito sessions ──
         deleted_empty = 0
         deleted_throwaway = 0
 
@@ -93,47 +84,13 @@ async def run_auto_sort(owner: str, skip_llm: bool = False, delete_throwaway: bo
             if is_session_recently_active(row, now=cleanup_now):
                 continue
 
-            msg_count = db.query(DbMsg.id).filter(
+            has_messages = db.query(DbMsg.id).filter(
                 DbMsg.session_id == row.id
-            ).limit(_THROWAWAY_MAX_MESSAGES + 1).count()
-            should_delete = False
-
-            if msg_count == 0:
+            ).first() is not None
+            if not has_messages:
                 if is_fresh:
                     continue
-                should_delete = True
                 deleted_empty += 1
-            elif delete_throwaway and msg_count <= _THROWAWAY_MAX_MESSAGES:
-                name = (row.name or "").strip().lower()
-                first_msg = db.query(DbMsg.content).filter(
-                    DbMsg.session_id == row.id, DbMsg.role == "user"
-                ).order_by(DbMsg.timestamp).first()
-                first_text = (first_msg[0] or "").strip().lower() if first_msg else ""
-                assistant_count = db.query(DbMsg.id).filter(
-                    DbMsg.session_id == row.id, DbMsg.role == "assistant"
-                ).limit(1).count()
-
-                if name in _THROWAWAY_NAMES or name.startswith("chat:") or first_text in _THROWAWAY_NAMES:
-                    should_delete = True
-                    deleted_throwaway += 1
-                elif msg_count == 1 and assistant_count == 0:
-                    should_delete = True
-                    deleted_throwaway += 1
-                elif msg_count <= 4 and first_text and len(first_text.split()) <= 8 and len(first_text) <= 80:
-                    # Short trivial chats — e.g. "write hi to a friend" → "Hi!"
-                    should_delete = True
-                    deleted_throwaway += 1
-                else:
-                    # Aggressive: total message text under 250 chars combined = trivial
-                    msg_rows = db.query(DbMsg.content).filter(
-                        DbMsg.session_id == row.id
-                    ).all()
-                    total_chars = sum(len(m[0] or "") for m in msg_rows)
-                    if total_chars <= 250:
-                        should_delete = True
-                        deleted_throwaway += 1
-
-            if should_delete:
                 db.delete(row)
 
         if deleted_empty or deleted_throwaway:
diff --git a/static/css/17-research.css b/static/css/17-research.css
index c4de5cb75..2c43aa2c8 100644
--- a/static/css/17-research.css
+++ b/static/css/17-research.css
@@ -2690,11 +2690,14 @@ body.theme-frosted .modal {
     max-height: 95dvh !important;
   }
 
+  /* Notes and Documents must respect the same viewport and safe-area
+     top as other mobile full-height sheets. Keep these authoritative rules
+     consistent with 05-documents.css and 14-email.css. */
   body.doc-view .doc-editor-pane,
   body.notes-view .notes-pane {
-    inset: auto 0 0 !important;
-    height: 95dvh !important;
-    max-height: 95dvh !important;
+    inset: var(--sheet-top) 0 0 !important;
+    height: calc(var(--app-height) - var(--sheet-top)) !important;
+    max-height: calc(var(--app-height) - var(--sheet-top)) !important;
   }
 }
 
diff --git a/static/index.html b/static/index.html
index aafacc2e1..df3045af3 100644
--- a/static/index.html
+++ b/static/index.html
@@ -16,13 +16,6 @@
   <link rel="apple-touch-icon" href="/static/icons/icon-192.png">
   <script nonce="{{CSP_NONCE}}">
   window._odysseusLoadTime = Date.now();
-  // Stops iOS focus-zoom on every field; iOS still allows pinch-zoom, other platforms would not.
-  (function(){
-    var ua = navigator.userAgent || '';
-    var isIOS = /iPad|iPhone|iPod/.test(ua) || (/Macintosh/.test(ua) && navigator.maxTouchPoints > 1);
-    var vp = document.querySelector('meta[name="viewport"]');
-    if (isIOS && vp) vp.setAttribute('content', vp.getAttribute('content') + ', maximum-scale=1');
-  })();
   (function(){
     try {
       var t = JSON.parse(localStorage.getItem('odysseus-theme'));
diff --git a/static/js/init.js b/static/js/init.js
index e554deaa7..516cbe186 100644
--- a/static/js/init.js
+++ b/static/js/init.js
@@ -379,7 +379,6 @@ document.addEventListener('DOMContentLoaded', markComposerUserEdited, { once: tr
       root.classList.toggle('keyboard-open', keyboardOpen);
       if (keyboardOpen) root.style.setProperty('--keyboard-viewport-height', vv.height + 'px');
       else root.style.removeProperty('--keyboard-viewport-height');
-      if (window.scrollY) window.scrollTo(0, 0);
     };
     // iOS fires visualViewport resize only once, mid-animation, and no focusin when
     // the tapped field already had focus (e.g. the composer after a new chat)
@@ -392,7 +391,6 @@ document.addEventListener('DOMContentLoaded', markComposerUserEdited, { once: tr
     window.visualViewport.addEventListener('scroll', _syncKeyboardViewport);
     document.addEventListener('focusin', _syncThroughKeyboardAnimation);
     document.addEventListener('focusout', _syncThroughKeyboardAnimation);
-    window.addEventListener('scroll', () => { if (window.scrollY) window.scrollTo(0, 0); }, { passive: true });
 
     let _lastVVHeight = window.visualViewport.height;
     window.visualViewport.addEventListener('resize', function() {
diff --git a/static/js/windowResize.js b/static/js/windowResize.js
index b71feac51..33469d3ed 100644
--- a/static/js/windowResize.js
+++ b/static/js/windowResize.js
@@ -267,8 +267,9 @@ export function makeWindowResizable(content, options = {}) {
           content.style.maxWidth = 'none';
           content.style.maxHeight = 'none';
           const left = parseFloat(content.style.left);
-          if (Number.isFinite(left) && left + w > window.innerWidth - padRight) {
-            content.style.left = Math.max(padLeft, window.innerWidth - padRight - w) + 'px';
+          if (Number.isFinite(left)) {
+            const rightmostLeft = Math.max(padLeft, window.innerWidth - padRight - w);
+            content.style.left = Math.max(padLeft, Math.min(left, rightmostLeft)) + 'px';
           }
         }
       } catch (_) {}
diff --git a/tests/TESTING_STANDARD.md b/tests/TESTING_STANDARD.md
index fd8c944e4..cb0eb71de 100644
--- a/tests/TESTING_STANDARD.md
+++ b/tests/TESTING_STANDARD.md
@@ -99,8 +99,10 @@ cleanup. Specifically:
   `os.environ[...] = ...` that outlives the test.
 - **Current working directory** - never `chdir` without restoring; never assert
   against cwd-relative paths like `./data`. Use a temp workspace helper instead.
-- **Database** - the root `conftest.py` defaults `DATABASE_URL` to an in-memory
-  SQLite for collection safety. A test that needs a real file-backed DB must opt
+- **Database** - the root `conftest.py` defaults `DATABASE_URL` to a private,
+  worker-owned SQLite file for collection safety and consistent schema across
+  TestClient threads. This fallback is not a per-test data fixture. A test that
+  writes or depends on particular rows must opt
   in explicitly via `tests.helpers.sqlite_db.make_temp_sqlite` and bind its
   `SessionLocal` onto the module under test. Do not rely on a persistent
   on-disk DB existing.
diff --git a/tests/conftest.py b/tests/conftest.py
index d6d60233c..ff5a354ea 100644
--- a/tests/conftest.py
+++ b/tests/conftest.py
@@ -12,9 +12,17 @@ sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
 # File-backed databases remain fixture-owned. Cleanup restores the caller.
 # Keep the existing generated environment-reference locations stable.
 _database_environment = pytest.MonkeyPatch()
-_database_environment.setenv("DATABASE_URL", "sqlite:///:memory:")
 from tests.helpers.worker_runtime import bootstrap_runtime, configure_runtime
 _runtime_environment = bootstrap_runtime()
+# A file keeps the collection schema consistent across TestClient threads;
+# SingletonThreadPool can evict the populated in-memory main-thread database.
+# One patcher owns this variable, including restoration of an inherited URL.
+_database_environment.setenv(
+    "DATABASE_URL",
+    f"sqlite:///{_runtime_environment.root / 'collection.db'}"
+    if _runtime_environment is not None else "sqlite:///:memory:",
+)
+_collection_database_engine = None
 
 # Pre-import real heavy modules BEFORE any test file's module-level stubs can
 # replace them with MagicMock. Some test files (e.g. test_llm_core_sanitize_*)
@@ -26,6 +34,7 @@ try:
     import sqlalchemy.orm  # noqa: F401
     import core.database  # noqa: F401
     import src.database
+    _collection_database_engine = core.database.engine
 except ImportError:
     pass  # not installed - the stubs below will handle it
 
@@ -100,6 +109,10 @@ def pytest_configure(config):
     only registers marker names; it imports no production module.
     """
     config.add_cleanup(_database_environment.undo)
+    if _collection_database_engine is not None:
+        # Registered after runtime cleanup, so connections close before the
+        # owned file is removed (also required on Windows).
+        config.add_cleanup(_collection_database_engine.dispose)
 
     import pathlib
     from tests._taxonomy import discover_markers
diff --git a/tests/e2e/README.md b/tests/e2e/README.md
index 859bf6b9d..9bd7c127f 100644
--- a/tests/e2e/README.md
+++ b/tests/e2e/README.md
@@ -1,5 +1,7 @@
 # Browser End-to-End Tests
 
+## Photo Editor Release Gate
+
 The photo editor release gate launches an isolated Odysseus server on port
 `7013` with a temporary SQLite database and disposable authenticated account.
 
@@ -18,3 +20,53 @@ The disposable account is created in the temporary data directory and exercises
 owner-scoped draft and notification routes without using an existing account.
 `PHOTO_EDITOR_E2E_AUTH=0` selects the legacy unauthenticated configuration for
 investigating it; offline-upgrade coverage requires the authenticated mode.
+
+## Mobile and iOS Layout Suite
+
+The dedicated mobile and iOS layout suite exercises viewport behavior,
+accessibility zoom, scrolling, mobile safe-area geometry (Notes and Documents
+panes), and floating window restoration against an isolated local Odysseus server
+running on port `17739`.
+
+```bash
+npm install
+npm run test:ios-layout:chromium
+```
+
+### Profiles and Browser Prerequisites
+
+- **Chromium profiles** (`desktop-chromium`, `phone-chromium-portrait`, `phone-chromium-landscape`, `tablet-chromium`):
+  Provide automated baseline coverage in CI and local development without requiring
+  WebKit dependencies.
+- **WebKit profiles** (`phone-webkit-portrait`, `phone-webkit-landscape`, `tablet-webkit`):
+  Used for platform-oriented WebKit rendering validation. WebKit requires browser binaries:
+
+```bash
+npm run test:ios-layout:install
+# or: npx playwright install webkit
+npm run test:ios-layout:webkit
+```
+
+To run all installed profiles:
+
+```bash
+npm run test:ios-layout
+```
+
+### Evidence and Cleanup
+
+- **Evidence artifacts**: Screenshots are written to `test-results/ios-layout/` by
+  default and traces are written to `test-results/ios-layout-traces/`. Both paths
+  are git-ignored. Set `ODYSSEUS_LAYOUT_EVIDENCE_DIR` to override the screenshot
+  output path.
+- **Teardown**: Disposable application state and database files created in the
+  OS temporary directory are automatically cleaned up on both normal and failure
+  teardown, preserving diagnostic screenshots and failure traces.
+
+### Platform Considerations
+
+Emulated Chromium and WebKit viewports validate deterministic CSS cascade,
+computed layout geometry, viewport meta tags, and JavaScript restoration. However,
+browser emulation is not a physical device substitute: physical iOS Safari and
+standalone installed PWAs should still be verified separately for native gesture
+pinch-zoom, software keyboard dismissal transitions, and dynamic hardware notch/island insets.
diff --git a/tests/e2e/ios-layout-teardown.cjs b/tests/e2e/ios-layout-teardown.cjs
new file mode 100644
index 000000000..5c42374dd
--- /dev/null
+++ b/tests/e2e/ios-layout-teardown.cjs
@@ -0,0 +1,28 @@
+const fs = require('node:fs');
+const os = require('node:os');
+const path = require('node:path');
+
+/**
+ * Global teardown for the iOS/mobile layout Playwright test suite.
+ * Cleans up only disposable temporary application data directories,
+ * preserving diagnostic artifacts (screenshots and failure traces) in test-results/.
+ */
+module.exports = async () => {
+  const dataDirectory = process.env.ODYSSEUS_LAYOUT_DATA_DIR;
+  if (!dataDirectory || typeof dataDirectory !== 'string') return;
+
+  try {
+    const resolvedData = path.resolve(dataDirectory);
+    const allowedTmpBases = [os.tmpdir(), '/var/tmp', '/tmp'].map(b => path.resolve(b));
+    const isTmp = allowedTmpBases.some(base => resolvedData.startsWith(base + path.sep));
+    const baseName = path.basename(resolvedData);
+
+    if (isTmp && (baseName.startsWith('odysseus-ios-layout-') || baseName.startsWith('odysseus-pr6396-'))) {
+      if (fs.existsSync(resolvedData)) {
+        fs.rmSync(resolvedData, { recursive: true, force: true });
+      }
+    }
+  } catch (err) {
+    console.warn(`[ios-layout-teardown] Warning: failed to clean temporary runtime dir: ${err.message}`);
+  }
+};
diff --git a/tests/e2e/ios-layout.config.cjs b/tests/e2e/ios-layout.config.cjs
new file mode 100644
index 000000000..73d2707ed
--- /dev/null
+++ b/tests/e2e/ios-layout.config.cjs
@@ -0,0 +1,148 @@
+// Tests against the running Odysseus app, not a static mock page.
+const fs = require('node:fs');
+const os = require('node:os');
+const path = require('node:path');
+const { defineConfig } = require('@playwright/test');
+
+const repositoryRoot = path.resolve(__dirname, '../..');
+const port = Number(process.env.ODYSSEUS_LAYOUT_PORT || '17739');
+const baseURL = `http://127.0.0.1:${port}`;
+const defaultPython = process.platform === 'win32'
+  ? path.join(repositoryRoot, '.venv', 'Scripts', 'python.exe')
+  : path.join(repositoryRoot, '.venv', 'bin', 'python');
+const python = process.env.ODYSSEUS_TEST_PYTHON
+  || (fs.existsSync(defaultPython) ? defaultPython : 'python3');
+
+const dataDirectory = process.env.ODYSSEUS_LAYOUT_DATA_DIR
+  || fs.mkdtempSync(path.join(os.tmpdir(), 'odysseus-ios-layout-'));
+process.env.ODYSSEUS_LAYOUT_DATA_DIR = dataDirectory;
+
+// Default evidence directory to repository-relative test-results/ios-layout
+// while preserving the environment-variable override if provided.
+const evidenceDir = process.env.ODYSSEUS_LAYOUT_EVIDENCE_DIR
+  || path.join(repositoryRoot, 'test-results', 'ios-layout');
+const outputDir = path.join(repositoryRoot, 'test-results', 'ios-layout-traces');
+
+// Emergency process-exit cleanup for disposable application runtime state
+const cleanDataDirSync = () => {
+  const dir = process.env.ODYSSEUS_LAYOUT_DATA_DIR;
+  if (!dir || typeof dir !== 'string') return;
+  try {
+    const resolved = path.resolve(dir);
+    const allowed = [os.tmpdir(), '/var/tmp', '/tmp'].map(b => path.resolve(b));
+    if (allowed.some(base => resolved.startsWith(base + path.sep))) {
+      const baseName = path.basename(resolved);
+      if (baseName.startsWith('odysseus-ios-layout-') || baseName.startsWith('odysseus-pr6396-')) {
+        if (fs.existsSync(resolved)) {
+          fs.rmSync(resolved, { recursive: true, force: true });
+        }
+      }
+    }
+  } catch (_) {}
+};
+
+process.once('SIGINT', () => { cleanDataDirSync(); process.exit(130); });
+process.once('SIGTERM', () => { cleanDataDirSync(); process.exit(143); });
+
+function isWebKitInstalled() {
+  try {
+    const { webkit } = require('@playwright/test');
+    return fs.existsSync(webkit.executablePath());
+  } catch (_) {
+    return false;
+  }
+}
+
+const chromiumProjects = [
+  {
+    name: 'desktop-chromium',
+    use: { browserName: 'chromium', viewport: { width: 1440, height: 900 } },
+  },
+  {
+    name: 'phone-chromium-portrait',
+    use: { browserName: 'chromium', viewport: { width: 390, height: 844 }, hasTouch: true, isMobile: true },
+  },
+  {
+    name: 'phone-chromium-landscape',
+    use: { browserName: 'chromium', viewport: { width: 844, height: 390 }, hasTouch: true, isMobile: true },
+  },
+  {
+    name: 'tablet-chromium',
+    use: { browserName: 'chromium', viewport: { width: 820, height: 1000 }, hasTouch: true, isMobile: true },
+  },
+];
+
+const webkitProjects = [
+  {
+    name: 'phone-webkit-portrait',
+    use: { browserName: 'webkit', viewport: { width: 390, height: 844 }, hasTouch: true, isMobile: true },
+  },
+  {
+    name: 'phone-webkit-landscape',
+    use: { browserName: 'webkit', viewport: { width: 844, height: 390 }, hasTouch: true, isMobile: true },
+  },
+  {
+    name: 'tablet-webkit',
+    use: { browserName: 'webkit', viewport: { width: 820, height: 1000 }, hasTouch: true, isMobile: true },
+  },
+];
+
+// Determine active projects based on environment and local browser availability
+const requestedBrowser = (process.env.ODYSSEUS_LAYOUT_BROWSER || 'all').toLowerCase();
+let selectedProjects = [];
+
+if (requestedBrowser === 'chromium') {
+  selectedProjects = chromiumProjects;
+} else if (requestedBrowser === 'webkit') {
+  selectedProjects = webkitProjects;
+} else {
+  // 'all': Include chromium always; include webkit if installed or if explicitly targeted via cli args
+  const hasCliWebKitTarget = process.argv.some(arg => arg.toLowerCase().includes('webkit'));
+  if (isWebKitInstalled() || hasCliWebKitTarget) {
+    selectedProjects = [...chromiumProjects, ...webkitProjects];
+  } else {
+    // Only Chromium is available locally (e.g. standard CI / default runner)
+    selectedProjects = chromiumProjects;
+  }
+}
+
+module.exports = defineConfig({
+  testDir: path.join(__dirname, 'ios-layout'),
+  testMatch: ['**/*.spec.cjs'],
+  outputDir,
+  timeout: 90_000,
+  expect: { timeout: 15_000 },
+  retries: 0,
+  workers: 1,
+  reporter: [['line']],
+  globalTeardown: require.resolve('./ios-layout-teardown.cjs'),
+  use: {
+    baseURL,
+    headless: true,
+    serviceWorkers: 'block',
+    screenshot: 'only-on-failure',
+    trace: 'retain-on-failure',
+  },
+  projects: selectedProjects,
+  metadata: { evidenceDir },
+  webServer: {
+    command: `${JSON.stringify(python)} -m uvicorn app:app --host 127.0.0.1 --port ${port}`,
+    cwd: repositoryRoot,
+    url: baseURL,
+    timeout: 120_000,
+    reuseExistingServer: false,
+    env: {
+      ...process.env,
+      AUTH_ENABLED: 'false',
+      // Never connect the disposable app to the host's real ChromaDB.
+      CHROMADB_HOST: '127.0.0.1',
+      CHROMADB_PORT: String(port + 1),
+      LLM_HOST: '127.0.0.1',
+      EMBEDDING_URL: `http://127.0.0.1:${port + 1}/v1/embeddings`,
+      DATABASE_URL: `sqlite:///${path.join(dataDirectory, 'app.db')}`,
+      ODYSSEUS_DATA_DIR: dataDirectory,
+      ODYSSEUS_STARTUP_WARMUPS: '0',
+      RESPONSE_COMPRESSION_ENABLED: 'false',
+    },
+  },
+});
diff --git a/tests/e2e/ios-layout/postmerge-layout.spec.cjs b/tests/e2e/ios-layout/postmerge-layout.spec.cjs
new file mode 100644
index 000000000..42f420032
--- /dev/null
+++ b/tests/e2e/ios-layout/postmerge-layout.spec.cjs
@@ -0,0 +1,313 @@
+const { test, expect } = require('@playwright/test');
+const fs = require('node:fs');
+const path = require('node:path');
+
+async function openApp(page) {
+  const response = await page.goto('/', { waitUntil: 'domcontentloaded' });
+  expect(response?.ok()).toBeTruthy();
+  // The runtime currently produces two #message nodes. Select the
+  // initialized composer while diagnosing whether both are visible.
+  const composers = await page.locator('textarea#message').evaluateAll(nodes =>
+    nodes.map((el, index) => ({
+      index,
+      initialized: el.dataset.startupPreserveBound === '1',
+      visible: !!el.getClientRects().length &&
+        getComputedStyle(el).visibility !== 'hidden',
+      parent: el.parentElement?.className || '',
+    }))
+  );
+  if (composers.length !== 1) {
+    console.log('Composer DOM diagnostic:', JSON.stringify(composers));
+  }
+  const activeComposer = page.locator(
+    'textarea#message[data-startup-preserve-bound="1"]'
+  );
+  await expect(activeComposer).toHaveCount(1);
+  await expect(activeComposer).toBeVisible();
+  expect(
+    composers.filter(composer => composer.visible).length,
+    'Exactly one composer should be visible'
+  ).toBe(1);
+}
+
+function isMobileViewport(testInfo) {
+  const width = testInfo.project.use?.viewport?.width || 0;
+  return width > 0 && width <= 768;
+}
+
+/**
+ * Reach Settings through genuine UI controls on either desktop or mobile viewports.
+ * On mobile (<=768px), the rail is hidden and the sidebar must be opened via hamburger.
+ */
+async function openSettings(page, isMobile) {
+  let settingsButton;
+  if (isMobile) {
+    const hamburger = page.locator('#hamburger-btn:visible').first();
+    await expect(hamburger, 'Hamburger button must be visible on mobile').toBeVisible({ timeout: 5000 });
+    await hamburger.click();
+    const sidebar = page.locator('#sidebar:not(.hidden)');
+    await expect(sidebar, 'Sidebar must open upon hamburger tap').toBeVisible({ timeout: 5000 });
+    await page.waitForTimeout(300);
+    settingsButton = sidebar.locator('#user-bar-settings');
+  } else {
+    settingsButton = page.locator('#rail-settings:visible, #user-bar-settings:visible').first();
+  }
+
+  await expect(settingsButton, 'Settings button must be reachable and visible').toBeVisible({ timeout: 5000 });
+  await settingsButton.click();
+  const dialog = page.locator('#settings-modal');
+  await expect(dialog, 'Settings modal must become visible').toBeVisible({ timeout: 5000 });
+  return dialog;
+}
+
+/**
+ * Reach Notes through genuine UI controls on either desktop or mobile viewports.
+ */
+async function openNotes(page, isMobile) {
+  let notesButton;
+  if (isMobile) {
+    const hamburger = page.locator('#hamburger-btn:visible').first();
+    await expect(hamburger, 'Hamburger button must be visible on mobile').toBeVisible({ timeout: 5000 });
+    await hamburger.click();
+    const sidebar = page.locator('#sidebar:not(.hidden)');
+    await expect(sidebar, 'Sidebar must open upon hamburger tap').toBeVisible({ timeout: 5000 });
+    await page.waitForTimeout(300);
+    notesButton = sidebar.locator('#tool-notes-btn');
+  } else {
+    notesButton = page.locator('#rail-notes:visible, #tool-notes-btn:visible').first();
+  }
+
+  await expect(notesButton, 'Notes button must be reachable and visible').toBeVisible({ timeout: 5000 });
+  await notesButton.click();
+  const realNotes = page.locator('.notes-pane').first();
+  await expect(realNotes, 'Notes pane must become visible').toBeVisible({ timeout: 5000 });
+  return realNotes;
+}
+
+/**
+ * Reach Document editor through genuine UI controls on either desktop or mobile viewports.
+ */
+async function openDocument(page, isMobile) {
+  let docButton;
+  if (isMobile) {
+    const hamburger = page.locator('#hamburger-btn:visible').first();
+    await expect(hamburger, 'Hamburger button must be visible on mobile').toBeVisible({ timeout: 5000 });
+    await hamburger.click();
+    const sidebar = page.locator('#sidebar:not(.hidden)');
+    await expect(sidebar, 'Sidebar must open upon hamburger tap').toBeVisible({ timeout: 5000 });
+    await page.waitForTimeout(300);
+    docButton = sidebar.locator('#library-new-doc-btn');
+  } else {
+    docButton = page.locator('#library-new-doc-btn:visible').first();
+    if (!(await docButton.count())) {
+      const hamburger = page.locator('#hamburger-btn:visible').first();
+      if (await hamburger.count()) {
+        await hamburger.click();
+        await page.waitForTimeout(300);
+        docButton = page.locator('#sidebar:not(.hidden) #library-new-doc-btn');
+      }
+    }
+  }
+
+  await expect(docButton, 'Document creation button must be reachable and visible').toBeVisible({ timeout: 5000 });
+  await docButton.click();
+  const docPane = page.locator('.doc-editor-pane').first();
+  await expect(docPane, 'Document editor pane must become visible').toBeVisible({ timeout: 10000 });
+  return docPane;
+}
+
+test('renders actual Odysseus UI and captures review screenshots', async ({ page }, testInfo) => {
+  await openApp(page);
+  const out = testInfo.config.metadata.evidenceDir;
+  fs.mkdirSync(out, { recursive: true });
+  const name = testInfo.project.name;
+  const isMobile = isMobileViewport(testInfo);
+
+  // 1. Chat screenshot (default view)
+  await page.screenshot({ path: path.join(out, `${name}-chat.png`), fullPage: false, animations: 'disabled' });
+
+  // 2. Settings screenshot via actual user-visible navigation
+  await openApp(page);
+  await openSettings(page, isMobile);
+  await page.screenshot({ path: path.join(out, `${name}-settings.png`), fullPage: false, animations: 'disabled' });
+
+  // 3. Notes screenshot via actual user-visible navigation
+  await openApp(page);
+  await openNotes(page, isMobile);
+  await page.screenshot({ path: path.join(out, `${name}-notes.png`), fullPage: false, animations: 'disabled' });
+
+  testInfo.annotations.push({ type: 'evidence', description: out });
+});
+
+test('viewport meta does not restrict user zoom', async ({ page }) => {
+  await openApp(page);
+  const meta = await page.locator('meta[name="viewport"]').getAttribute('content');
+  expect(meta || '').not.toMatch(/maximum-scale\s*=\s*1(?:\b|\.)/i);
+  expect(meta || '').not.toMatch(/user-scalable\s*=\s*no/i);
+});
+
+test('browser-owned window scrolling is not globally reset', async ({ page }) => {
+  await openApp(page);
+  const value = await page.evaluate(async () => {
+    document.documentElement.style.overflowY = 'auto';
+    document.body.style.overflowY = 'auto';
+    document.body.style.height = 'auto';
+    document.body.style.minHeight = '3200px';
+    const spacer = document.createElement('div');
+    spacer.style.cssText = 'position:absolute;top:3100px;width:1px;height:10px;pointer-events:none;';
+    document.body.appendChild(spacer);
+    window.scrollTo(0, 250);
+    await new Promise(resolve => setTimeout(resolve, 150));
+    return window.scrollY;
+  });
+  expect(value).toBeGreaterThan(100);
+});
+
+test('CSS cascade rules for mobile Notes and Documents respect safe areas', async ({ page }, testInfo) => {
+  test.skip(!isMobileViewport(testInfo), 'Only mobile viewports <=768px');
+  await openApp(page);
+  const results = await page.evaluate(() => {
+    document.documentElement.style.setProperty('--safe-top', '44px');
+    document.documentElement.style.setProperty('--sheet-top', '68px');
+    document.documentElement.style.setProperty('--app-height', '844px');
+    const values = {};
+    for (const [key, bodyClass, paneClass] of [
+      ['notes', 'notes-view', 'notes-pane'],
+      ['document', 'doc-view', 'doc-editor-pane'],
+    ]) {
+      const pane = document.createElement('section');
+      pane.className = paneClass;
+      pane.style.opacity = '0'; // CSS cascade regression check
+      document.body.appendChild(pane);
+      document.body.classList.add(bodyClass);
+      const style = window.getComputedStyle(pane);
+      values[key] = {
+        top: style.top,
+        height: style.height,
+        insetTop: pane.getBoundingClientRect().top,
+      };
+      pane.remove();
+      document.body.classList.remove(bodyClass);
+    }
+    return values;
+  });
+  for (const key of ['notes', 'document']) {
+    expect(results[key].top, key).toBe('68px');
+    expect(parseFloat(results[key].height), key).toBeGreaterThanOrEqual(770);
+  }
+});
+
+test('real Notes and Document panes honor safe-area geometry in the live application', async ({ page }, testInfo) => {
+  const isMobile = isMobileViewport(testInfo);
+  test.skip(!isMobile, 'Only mobile viewports <=768px');
+  await openApp(page);
+
+  // Configure simulated safe-area insets (e.g. iPhone notch / Dynamic Island)
+  await page.evaluate(() => {
+    document.documentElement.style.setProperty('--safe-top', '44px');
+    document.documentElement.style.setProperty('--sheet-top', '68px');
+    document.documentElement.style.setProperty('--app-height', '844px');
+  });
+
+  // 1. Verify actual Notes UI
+  const realNotes = await openNotes(page, isMobile);
+  const notesMetrics = await realNotes.evaluate(el => {
+    const style = window.getComputedStyle(el);
+    const rect = el.getBoundingClientRect();
+    return {
+      top: style.top,
+      height: style.height,
+      maxHeight: style.maxHeight,
+      rectTop: rect.top,
+      rectHeight: rect.height,
+    };
+  });
+
+  expect(notesMetrics.top, 'Real Notes pane top must match --sheet-top').toBe('68px');
+  expect(parseFloat(notesMetrics.height), 'Real Notes pane height must respect safe area').toBeGreaterThanOrEqual(770);
+  expect(notesMetrics.rectTop, 'Real Notes pane must not overlap the 44px safe-area top notch').toBeGreaterThanOrEqual(44);
+
+  // 2. Verify actual Document UI
+  await openApp(page);
+  await page.evaluate(() => {
+    document.documentElement.style.setProperty('--safe-top', '44px');
+    document.documentElement.style.setProperty('--sheet-top', '68px');
+    document.documentElement.style.setProperty('--app-height', '844px');
+  });
+
+  const realDoc = await openDocument(page, isMobile);
+  const docMetrics = await realDoc.evaluate(el => {
+    const style = window.getComputedStyle(el);
+    const rect = el.getBoundingClientRect();
+    return {
+      top: style.top,
+      height: style.height,
+      maxHeight: style.maxHeight,
+      rectTop: rect.top,
+      rectHeight: rect.height,
+    };
+  });
+
+  expect(docMetrics.top, 'Real Document pane top must match --sheet-top').toBe('68px');
+  expect(parseFloat(docMetrics.height), 'Real Document pane height must respect safe area').toBeGreaterThanOrEqual(770);
+  expect(docMetrics.rectTop, 'Real Document pane must not overlap the 44px safe-area top notch').toBeGreaterThanOrEqual(44);
+});
+
+test('restored floating windows clamp away from a left safe area', async ({ page }, testInfo) => {
+  test.skip(isMobileViewport(testInfo), 'Geometry test uses desktop resizable windows');
+  await openApp(page);
+  const value = await page.evaluate(async () => {
+    document.documentElement.style.setProperty('--safe-left', '58px');
+    document.body.style.setProperty('padding-left', '58px', 'important');
+    const probe = document.createElement('div');
+    probe.style.cssText = 'position:fixed;left:10px;top:80px;width:420px;height:250px;z-index:999999;';
+    document.body.appendChild(probe);
+    const key = 'odysseus-layout-restore-test';
+    localStorage.setItem(key, JSON.stringify({ w: 480, h: 280 }));
+    try {
+      // Body horizontal padding transitions over 160ms. Wait for the
+      // simulated safe-area change to settle before restoring geometry.
+      await new Promise(resolve => setTimeout(resolve, 300));
+
+      const padAtRestore = parseFloat(
+        getComputedStyle(document.body).paddingLeft
+      );
+
+      if (Math.abs(padAtRestore - 58) > 0.5) {
+        throw new Error(
+          `Safe-area fixture did not settle: ${padAtRestore}px`
+        );
+      }
+
+      // Verify the running server actually serves the patched module.
+      const response = await fetch(
+        '/static/js/windowResize.js',
+        { cache: 'no-store' }
+      );
+      if (!response.ok) {
+        throw new Error('Cannot fetch windowResize.js');
+      }
+      const moduleSource = await response.text();
+      if (!moduleSource.includes('const rightmostLeft = Math.max(padLeft')) {
+        throw new Error('Running server is serving unpatched windowResize.js');
+      }
+
+      const { makeWindowResizable } = await import('/static/js/windowResize.js');
+      makeWindowResizable(probe, { storageKey: key, mobileSkip: 768 });
+      await new Promise(resolve => setTimeout(resolve, 300));
+      return {
+        left: parseFloat(probe.style.left),
+        width: parseFloat(probe.style.width),
+        paddingLeft: parseFloat(getComputedStyle(document.body).paddingLeft),
+        viewportWidth: window.innerWidth,
+      };
+    } finally {
+      localStorage.removeItem(key);
+      probe.remove();
+    }
+  });
+
+  expect(value.paddingLeft).toBeGreaterThanOrEqual(58);
+  expect(value.width).toBe(480);
+  expect(value.left).toBeGreaterThanOrEqual(value.paddingLeft);
+});
diff --git a/tests/test_admin_listing_authorization.py b/tests/test_admin_listing_authorization.py
index 1fd3c4c84..db6c36cfe 100644
--- a/tests/test_admin_listing_authorization.py
+++ b/tests/test_admin_listing_authorization.py
@@ -14,12 +14,22 @@ from routes.auth_routes import SESSION_COOKIE
 from tests.helpers.database import disposable_database
 
 
-@pytest.fixture
-def authenticated_app(tmp_path, monkeypatch):
-    monkeypatch.setenv("AUTH_ENABLED", "true")
-    auth = AuthManager(str(tmp_path / "auth.json"))
+@pytest.fixture(scope="module")
+def auth_configuration(tmp_path_factory):
+    """Hash real fixture passwords once; only immutable bytes are reused."""
+    path = tmp_path_factory.mktemp("admin-auth-seed") / "auth.json"
+    auth = AuthManager(str(path))
     for user, admin in (("member", False), ("admin", True)):
         assert auth.create_user(user, "disposable-test-password", is_admin=admin)
+    return path.read_bytes()
+
+
+@pytest.fixture
+def authenticated_app(tmp_path, monkeypatch, auth_configuration):
+    monkeypatch.setenv("AUTH_ENABLED", "true")
+    path = tmp_path / "auth.json"
+    path.write_bytes(auth_configuration)
+    auth = AuthManager(str(path))
     app = FastAPI()
     app.state.auth_manager = auth
 
@@ -133,3 +143,17 @@ def test_diagnostics_require_real_admin_before_sensitive_work(
         for operation in (stats, rag.get_stats, research.call_research_service,
                           extract_id, transcript, collect, read_log):
             operation.assert_not_called()
+
+
+@pytest.mark.parametrize("instance", [0, 1])
+def test_authenticated_fixture_has_private_users_sessions_and_file(authenticated_app, instance):
+    app, client_for = authenticated_app
+    auth = app.state.auth_manager
+    assert set(auth.users) == {"member", "admin"}
+    assert not auth._sessions
+    with client_for("member") as client:
+        assert auth.get_username_for_token(client.cookies.get(SESSION_COOKIE)) == "member"
+    assert auth.create_user("fixture-local", "disposable-test-password")
+    # Both parameter cases mutate their own file. Any shared manager, mutable
+    # users/sessions or written-back seed makes the next case fail, in any order.
+    assert "fixture-local" in auth.users
diff --git a/tests/test_agent_editor_routes.py b/tests/test_agent_editor_routes.py
index f9f27e3e1..89641276b 100644
--- a/tests/test_agent_editor_routes.py
+++ b/tests/test_agent_editor_routes.py
@@ -1929,23 +1929,60 @@ def test_place_layer_reports_false_on_queued_response(monkeypatch):
     assert placed is False
 
 
-def test_bridge_await_cancellation_handled_cleanly():
+def test_bridge_await_cancellation_handled_cleanly(monkeypatch):
     import asyncio
+    from contextlib import suppress
     from httpx import ASGITransport, AsyncClient
     app = _bridge_app()
 
     async def scenario():
+        entered_wait = asyncio.Event()
+        real_wait_for = asyncio.wait_for
+        waits = []
+
+        async def observe_wait(awaitable, timeout):
+            # Observe the real wait; do not replace its result or cancellation.
+            waits.append(awaitable)
+            entered_wait.set()
+            return await real_wait_for(awaitable, timeout)
+
+        monkeypatch.setattr(eaar.asyncio, "wait_for", observe_wait)
         async with AsyncClient(transport=ASGITransport(app=app), base_url="http://test") as c:
             cmd_resp = await c.post("/api/editor/bridge/command", json={"session": "cancel-sess", "op": "document.info", "wait": False})
+            assert cmd_resp.status_code == 200
+            assert cmd_resp.json()["completed"] is False
             cid = cmd_resp.json()["queued"][0]
 
             await_task = asyncio.create_task(c.post("/api/editor/bridge/await", json={"id": cid, "timeout": 30}))
-            await asyncio.sleep(0.05)
-            await_task.cancel()
             try:
-                await await_task
-            except asyncio.CancelledError:
-                pass
+                await real_wait_for(entered_wait.wait(), 5)
+                assert len(waits) == 1
+                assert not waits[0].done()
+                assert not await_task.done(), "await returned without a queued result"
+                await_task.cancel()
+                response = await real_wait_for(await_task, 5)
+                assert response.status_code == 200
+                assert response.json() == {
+                    "ok": False, "error": "cancelled", "status": "cancelled",
+                    "completed": False, "id": cid,
+                }
+            finally:
+                if not await_task.done():
+                    await_task.cancel()
+                    with suppress(asyncio.CancelledError):
+                        await await_task
+
+            # Cancelling the waiter must not cancel the queued command itself.
+            payload = {"width": 137, "height": 251}
+            delivered = await c.post("/api/editor/bridge/result", json={"id": cid, "ok": True, "result": payload})
+            assert delivered.status_code == 200
+            completed = await c.post("/api/editor/bridge/await", json={"id": cid, "timeout": 1})
+            assert completed.status_code == 200
+            assert completed.json()["status"] == "completed"
+            assert completed.json()["completed"] is True
+            assert completed.json()["ok"] is True
+            assert completed.json()["result"] == payload
+            assert len(waits) == 2
 
     asyncio.run(scenario())
 
diff --git a/tests/test_auth_config_lock_concurrency.py b/tests/test_auth_config_lock_concurrency.py
index 34232b9e2..d1707e5b4 100644
--- a/tests/test_auth_config_lock_concurrency.py
+++ b/tests/test_auth_config_lock_concurrency.py
@@ -206,35 +206,50 @@ class TestDiskConsistency:
     @pytest.mark.slow
     def test_file_always_valid_json_during_concurrent_ops(self, tmp_path):
         mgr = _fresh_auth_manager(tmp_path)
-        mgr.create_user("admin", "adminpw", is_admin=True)
+        assert mgr.create_user("admin", "adminpw", is_admin=True)
 
         stop_event = threading.Event()
+        reader_ready = threading.Event()
         corruption_found = []
+        observations = {"reads": 0, "with_writes": 0}
 
         def reader():
             while not stop_event.is_set():
                 try:
                     with open(mgr.auth_path, "r") as f:
                         content = f.read()
-                    json.loads(content)
+                    data = json.loads(content)
+                    observations["reads"] += 1
+                    if any(name.startswith("stress") for name in data["users"]):
+                        observations["with_writes"] += 1
+                    reader_ready.set()
                 except json.JSONDecodeError as e:
                     corruption_found.append(str(e))
                     break
-                except FileNotFoundError:
-                    pass
                 time.sleep(0.001)
 
         def writer():
+            assert reader_ready.wait(5), "reader never observed the initial file"
+            results = []
             for i in range(50):
-                mgr.create_user(f"stress{i}", f"pw{i}")
+                results.append(mgr.create_user(f"stress{i}", f"pw{i}"))
+            return results
 
-        reader_thread = threading.Thread(target=reader)
-        writer_thread = threading.Thread(target=writer)
-
-        reader_thread.start()
-        writer_thread.start()
-        writer_thread.join()
-        stop_event.set()
-        reader_thread.join()
+        # Futures propagate reader/writer exceptions to the test thread.
+        with ThreadPoolExecutor(max_workers=2) as pool:
+            reading = pool.submit(reader)
+            writing = pool.submit(writer)
+            try:
+                results = writing.result(timeout=120)
+            finally:
+                stop_event.set()
+            reading.result(timeout=5)
 
         assert not corruption_found, f"Corrupt JSON detected: {corruption_found[0]}"
+        assert len(results) == 50 and all(results), "writer operations did not succeed"
+        assert observations["reads"] > 0
+        assert observations["with_writes"] > 0, "reader saw no persisted writer activity"
+        expected = {"admin"} | {f"stress{i}" for i in range(50)}
+        with open(mgr.auth_path) as persisted:
+            assert set(json.load(persisted)["users"]) == expected
+        assert set(_fresh_auth_manager(tmp_path).users) == expected
diff --git a/tests/test_caldav_redirect_hardening.py b/tests/test_caldav_redirect_hardening.py
index cd82974aa..6a0f76345 100644
--- a/tests/test_caldav_redirect_hardening.py
+++ b/tests/test_caldav_redirect_hardening.py
@@ -13,6 +13,7 @@ import socketserver
 import threading
 
 import pytest
+import requests
 
 from src import caldav_sync, caldav_writeback
 
@@ -26,12 +27,13 @@ def test_build_dav_client_disables_redirects():
 
 def test_dav_client_does_not_follow_redirect_to_internal_host():
     """End-to-end through the real DAVClient: a 302 toward an internal host
-    must NOT be followed. Without the fix the sink is contacted (SSRF); with it
-    the redirect is refused and the sink is never reached."""
+    must NOT be followed. Exercise the session limit separately because some
+    caldav versions independently refuse cross-origin PROPFIND redirects."""
     pytest.importorskip("caldav")
 
     sink_hits: list[str] = []
     public_methods: list[str] = []
+    same_origin_hits: list[str] = []
 
     class _Internal(http.server.BaseHTTPRequestHandler):
         # Stand-in for an internal service the attacker redirects toward.
@@ -48,9 +50,15 @@ def test_dav_client_does_not_follow_redirect_to_internal_host():
     class _Public(http.server.BaseHTTPRequestHandler):
         # The "validated" public CalDAV server that redirects everything inward.
         def do_GET(self):  # noqa: N802
+            if self.path == "/landing":
+                same_origin_hits.append(self.path)
+                self.send_response(200)
+                self.end_headers()
+                return
             public_methods.append(self.command)
             self.send_response(302)
-            self.send_header("Location", f"http://127.0.0.1:{internal_port}/leak")
+            location = "/landing" if self.path == "/same" else f"http://127.0.0.1:{internal_port}/leak"
+            self.send_header("Location", location)
             self.end_headers()
 
         do_PROPFIND = do_GET
@@ -62,8 +70,10 @@ def test_dav_client_does_not_follow_redirect_to_internal_host():
     internal_port = internal.server_address[1]
     public = socketserver.TCPServer(("127.0.0.1", 0), _Public)
     public_port = public.server_address[1]
-    threading.Thread(target=internal.serve_forever, daemon=True).start()
-    threading.Thread(target=public.serve_forever, daemon=True).start()
+    internal_thread = threading.Thread(target=internal.serve_forever, daemon=True)
+    public_thread = threading.Thread(target=public.serve_forever, daemon=True)
+    internal_thread.start()
+    public_thread.start()
     try:
         public_url = f"http://127.0.0.1:{public_port}/dav"
         client = caldav_sync._build_dav_client(public_url, "u", "p")
@@ -82,17 +92,35 @@ def test_dav_client_does_not_follow_redirect_to_internal_host():
         assert public_methods in (["PROPFIND"], ["PROPFIND", "GET"]), public_methods
         assert sink_hits == [], "redirect toward an internal host must not be followed"
 
-        # Positive control: demonstrate that the sink would see the redirect
-        # if the session's redirect protection were removed.
+        # GET redirects are handled by the HTTP session even in caldav 3.7.
+        # This makes the cross-port SSRF assertion sensitive to our own limit.
+        with pytest.raises(Exception, match="Exceeded 0 redirects"):
+            client.request(public_url, "GET", "")
+        assert sink_hits == [], "GET redirect toward the internal sink must be blocked"
+
+        # requests can reach the cross-port sink when redirects are enabled,
+        # even if DAVClient independently rejects cross-origin PROPFIND redirects.
+        with requests.Session() as session:
+            response = session.get(public_url, timeout=5)
+            assert response.status_code == 207
+        assert sink_hits == ["/leak"], "the cross-port redirect must reach the sink"
+
+        # GET follows redirects at the HTTP session layer across these caldav
+        # versions. Check the application's limit on a same-origin redirect too.
+        same_url = f"http://127.0.0.1:{public_port}/same"
+        with pytest.raises(Exception, match="Exceeded 0 redirects"):
+            client.request(same_url, "GET", "")
+        assert same_origin_hits == []
         client.session.max_redirects = 30
-        try:
-            client.request(public_url, "PROPFIND", "")
-        except Exception:
-            pass  # The sink's response is intentionally not a valid DAV body.
-        assert sink_hits == ["/leak"], "the test must detect a followed redirect"
+        client.request(same_url, "GET", "")
+        assert same_origin_hits == ["/landing"]
     finally:
         internal.shutdown()
         public.shutdown()
+        internal_thread.join()
+        public_thread.join()
+        internal.server_close()
+        public.server_close()
 
 
 def test_sync_and_writeback_construct_clients_through_the_helper():
diff --git a/tests/test_chat_tidy_content_preservation.py b/tests/test_chat_tidy_content_preservation.py
new file mode 100644
index 000000000..b203eac5f
--- /dev/null
+++ b/tests/test_chat_tidy_content_preservation.py
@@ -0,0 +1,148 @@
+"""Both chat tidy entry points retain saved conversations and still remove empties."""
+import asyncio
+from datetime import datetime, timedelta, timezone
+
+import pytest
+from fastapi import APIRouter, FastAPI
+from fastapi.testclient import TestClient
+from sqlalchemy import create_engine
+from sqlalchemy.orm import sessionmaker
+
+import core.database as database
+import core.session_manager as manager_module
+import routes.session_routes as routes
+import src.session_actions as actions
+
+
+@pytest.fixture
+def tidy_api(tmp_path, monkeypatch):
+    engine = create_engine(
+        f"sqlite:///{(tmp_path / 'tidy.db').as_posix()}",
+        connect_args={"check_same_thread": False},
+    )
+    database.Base.metadata.create_all(engine)
+    factory = sessionmaker(bind=engine, autoflush=False)
+    monkeypatch.setattr(database, "SessionLocal", factory)
+    monkeypatch.setattr(manager_module, "SessionLocal", factory)
+    monkeypatch.setattr(routes, "SessionLocal", factory)
+    monkeypatch.setattr(routes, "router", APIRouter(prefix="/api"))
+    monkeypatch.setattr(routes, "effective_user", lambda request: "alice")
+    manager = manager_module.SessionManager.__new__(manager_module.SessionManager)
+    manager.sessions = {}
+    manager.upload_handler = None
+    monkeypatch.setattr(manager, "get_sessions_for_user", lambda owner: {})
+    app = FastAPI()
+    app.include_router(routes.setup_session_routes(manager, {}))
+    with TestClient(app) as client:
+        yield client, factory, manager
+    engine.dispose()
+
+
+def run_tidy(client, entrypoint):
+    if entrypoint == "action":
+        return asyncio.run(actions.run_auto_sort("alice", skip_llm=True))
+    response = client.post("/api/sessions/auto-sort?skip_llm=true")
+    assert response.status_code == 200, response.text
+    return response.json()
+
+
+@pytest.mark.parametrize("entrypoint", ["route", "action"])
+@pytest.mark.parametrize("title,prompt", [
+    ("New chat", "Please analyze the following contract and retain all details. " * 20),
+    ("Contract analysis", "Review contract"),
+    ("Chat: analysis", "Explain these results and retain the full answer"),
+    ("Test", "Here is a large input that should be analyzed and retained"),
+], ids=["generic-title", "short-prompt", "title-prefix", "test-title"])
+def test_tidy_preserves_nontrivial_saved_answer(tidy_api, title, prompt, entrypoint):
+    client, factory, manager = tidy_api
+    old = datetime.now(timezone.utc).replace(tzinfo=None) - timedelta(hours=2)
+    answer = "Clause analysis and recommended changes.\n" * 2000
+    with factory() as db:
+        db.add(database.Session(
+            id="valuable-chat", name=title, owner="alice", endpoint_url="", model="",
+            archived=False, is_important=False, message_count=2,
+            created_at=old, updated_at=old, last_accessed=old, last_message_at=old,
+        ))
+        db.commit()
+        db.add_all([
+            database.ChatMessage(id="prompt", session_id="valuable-chat", role="user",
+                                 content=prompt, timestamp=old),
+            database.ChatMessage(id="answer", session_id="valuable-chat", role="assistant",
+                                 content=answer, timestamp=old + timedelta(seconds=1)),
+        ])
+        db.commit()
+    run_tidy(client, entrypoint)
+    with factory() as db:
+        remaining = db.query(database.ChatMessage).filter_by(session_id="valuable-chat").count()
+        survives = db.get(database.Session, "valuable-chat") is not None
+    assert survives and remaining == 2, "Title/prompt heuristics must not erase saved answers"
+    with factory() as db:
+        assert db.get(database.ChatMessage, "answer").content == answer
+
+
+@pytest.mark.parametrize("entrypoint", ["route", "action"])
+@pytest.mark.parametrize("kind", ["unanswered", "greeting", "empty", "fresh-empty", "starred-empty", "archived", "other-owner", "incognito"])
+def test_tidy_cleanup_boundaries(tidy_api, entrypoint, kind):
+    client, factory, manager = tidy_api
+    now = datetime.now(timezone.utc).replace(tzinfo=None)
+    timestamp = now if kind == "fresh-empty" else now - timedelta(hours=2)
+    message = kind in {"unanswered", "greeting", "archived", "other-owner", "incognito"}
+    with factory() as db:
+        db.add(database.Session(
+            id="boundary-chat", name="Incognito" if kind == "incognito" else "New chat",
+            owner="bob" if kind == "other-owner" else "alice", endpoint_url="", model="",
+            archived=kind == "archived", is_important=kind == "starred-empty",
+            # Deliberately inaccurate metadata: cleanup must use message rows.
+            message_count=0 if message else 40,
+            created_at=timestamp, updated_at=timestamp, last_accessed=timestamp, last_message_at=timestamp,
+        ))
+        db.commit()
+        if message:
+            db.add(database.ChatMessage(id="boundary-message", session_id="boundary-chat", role="user",
+                                        content="hi" if kind == "greeting" else "Pending question " * 100,
+                                        timestamp=timestamp))
+            db.commit()
+    run_tidy(client, entrypoint)
+    deleted = kind in {"empty", "incognito"}
+    with factory() as db:
+        assert (db.get(database.Session, "boundary-chat") is None) == deleted
+        assert db.query(database.ChatMessage).filter_by(session_id="boundary-chat").count() == int(message and not deleted)
+
+
+@pytest.mark.parametrize("entrypoint", ["route", "action"])
+def test_tidy_keeps_folder_assignment_working(tidy_api, monkeypatch, entrypoint):
+    from types import SimpleNamespace
+    import src.llm_core as llm
+    import src.task_endpoint as endpoints
+    client, factory, manager = tidy_api
+    old = datetime.now(timezone.utc).replace(tzinfo=None) - timedelta(hours=2)
+    for sid in ["chat-one-long-id", "chat-two-long-id"]:
+        with factory() as db:
+            db.add(database.Session(id=sid, name="New chat", owner="alice", endpoint_url="", model="",
+                                    archived=False, created_at=old, updated_at=old,
+                                    last_accessed=old, last_message_at=old))
+            db.commit()
+            db.add(database.ChatMessage(id=f"{sid}-message", session_id=sid, role="user",
+                                        content="Saved question", timestamp=old))
+            db.commit()
+        manager.sessions[sid] = SimpleNamespace(id=sid, name="New chat", archived=False,
+                                                updated_at=old.isoformat(), created_at=old.isoformat())
+    monkeypatch.setattr(manager, "get_sessions_for_user", lambda owner: manager.sessions)
+    monkeypatch.setattr(endpoints, "resolve_task_endpoint", lambda *args, **kwargs: ("http://model.invalid", "test", {}))
+    response_text = '{"folders":{"Projects":["chat-one","chat-two"]}}'
+    monkeypatch.setattr(llm, "llm_call", lambda *args, **kwargs: response_text)
+
+    async def async_response(*args, **kwargs):
+        return response_text
+
+    monkeypatch.setattr(llm, "llm_call_async", async_response)
+    if entrypoint == "action":
+        result = asyncio.run(actions.run_auto_sort("alice"))
+        assert "Sorted 2 sessions" in result
+    else:
+        response = client.post("/api/sessions/auto-sort")
+        assert response.status_code == 200, response.text
+        assert response.json()["updated"] == 2
+    with factory() as db:
+        assert db.query(database.ChatMessage).count() == 2
+        assert {row.folder for row in db.query(database.Session).all()} == {"Projects"}
diff --git a/tests/test_cookbook_dependency_completion_regression.py b/tests/test_cookbook_dependency_completion_regression.py
index 60e7f2b53..2366e5098 100644
--- a/tests/test_cookbook_dependency_completion_regression.py
+++ b/tests/test_cookbook_dependency_completion_regression.py
@@ -1,4 +1,14 @@
+import os
+import subprocess
 from pathlib import Path
+import pytest
+from core.platform_compat import find_bash, git_bash_path
+from routes.cookbook_helpers import (
+    ODYSSEUS_LLAMA_SHIM_MARKER,
+    _append_llama_server_shim_definition_lines,
+    _append_llama_server_shim_reconciliation_lines,
+    _append_llama_server_finalization_lines,
+)
 
 from tests.helpers.stylesheets import app_css
 
@@ -615,3 +625,730 @@ def test_local_dependency_probe_refreshes_user_site_visibility():
     # deps (realesrgan) probe as not-installed until a restart. See #4810.
     assert "if user_site and os.path.isdir(user_site):" in source
     assert "site.addsitedir(user_site)" in source
+
+
+def test_llama_runner_path_preserves_standard_user_bin_order():
+    """The runner PATH must preserve standard user $HOME/bin precedence while
+    relying on ownership-aware reconciliation to retire Odysseus shims when
+    native builds exist."""
+    source = _read("routes/cookbook_routes.py")
+
+    assert (
+        'export PATH="$HOME/.local/bin:$HOME/bin:$HOME/llama.cpp/build/bin:/opt/homebrew/bin:/usr/local/bin:$PATH"'
+        in source
+    )
+
+
+def _run_shell(tmp_path, lines, cwd=None, path=None, extra_env=None):
+    from core.platform_compat import find_bash, git_bash_path
+    bash = find_bash() or "bash"
+    env = dict(os.environ, HOME=git_bash_path(tmp_path))
+    if path is not None:
+        env["PATH"] = path
+    if extra_env:
+        env.update(extra_env)
+    return subprocess.run(
+        [bash, "-c", "\n".join(lines)],
+        cwd=str(cwd) if cwd else None,
+        env=env,
+        capture_output=True,
+        text=True,
+    )
+
+
+def _write_odysseus_shim(path: Path) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text(
+        f"#!/usr/bin/env bash\n{ODYSSEUS_LLAMA_SHIM_MARKER}\nexec python3 -m llama_cpp.server \"$@\"\n",
+        encoding="utf-8",
+    )
+    path.chmod(0o755)
+
+
+def _write_mock_native(path: Path, succeeds: bool = True) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    if succeeds:
+        content = '#!/usr/bin/env bash\nif [ "$1" = "--version" ]; then echo "version: 0.1.0-dev"; exit 0; fi\necho "RUN_NATIVE: $@"\nexit 0\n'
+    else:
+        content = '#!/usr/bin/env bash\necho "error: missing libllama.so" >&2\nexit 127\n'
+    path.write_text(content, encoding="utf-8")
+    path.chmod(0o755)
+
+
+def _write_mock_python3(path: Path, succeeds: bool = True) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text(f"#!/usr/bin/env bash\nexit {0 if succeeds else 1}\n", encoding="utf-8")
+    path.chmod(0o755)
+
+
+def test_odysseus_create_llama_shim_never_overwrites_user_regular_file(tmp_path):
+    """Creation boundary: never overwrite a non-executable user-owned regular file."""
+    user_file = tmp_path / "bin/llama-server"
+    user_file.parent.mkdir(parents=True, exist_ok=True)
+    original_content = "user configuration or binary data without marker"
+    user_file.write_text(original_content, encoding="utf-8")
+
+    lines = []
+    _append_llama_server_shim_definition_lines(lines)
+    lines.append('_odysseus_create_llama_shim')
+
+    res = _run_shell(tmp_path, lines)
+    assert res.returncode != 0, "Creation should return non-zero when target is a user file"
+    assert user_file.read_text(encoding="utf-8") == original_content, "User file must remain byte-for-byte unchanged"
+
+
+def test_odysseus_create_llama_shim_never_overwrites_user_symlink(tmp_path):
+    """Creation boundary: never follow or overwrite a user-owned symlink."""
+    bin_dir = tmp_path / "bin"
+    bin_dir.mkdir(parents=True, exist_ok=True)
+    target_file = tmp_path / "user_target.txt"
+    target_file.write_text("precious user target content", encoding="utf-8")
+    user_symlink = bin_dir / "llama-server"
+    user_symlink.symlink_to(target_file)
+
+    lines = []
+    _append_llama_server_shim_definition_lines(lines)
+    lines.append('_odysseus_create_llama_shim')
+
+    res = _run_shell(tmp_path, lines)
+    assert res.returncode != 0, "Creation should return non-zero when target is a user symlink"
+    assert user_symlink.is_symlink(), "Symlink must remain intact"
+    assert target_file.read_text(encoding="utf-8") == "precious user target content", "Target file must not be modified"
+
+
+def test_odysseus_create_llama_shim_never_overwrites_dangling_symlink(tmp_path):
+    """Creation boundary: never overwrite a dangling user-owned symlink."""
+    bin_dir = tmp_path / "bin"
+    bin_dir.mkdir(parents=True, exist_ok=True)
+    dangling_symlink = bin_dir / "llama-server"
+    dangling_symlink.symlink_to(tmp_path / "nonexistent_target")
+
+    lines = []
+    _append_llama_server_shim_definition_lines(lines)
+    lines.append('_odysseus_create_llama_shim')
+
+    res = _run_shell(tmp_path, lines)
+    assert res.returncode != 0, "Creation should return non-zero when target is a dangling symlink"
+    assert dangling_symlink.is_symlink(), "Dangling symlink must remain intact"
+
+
+def test_odysseus_create_llama_shim_safely_replaces_existing_odysseus_shim(tmp_path):
+    """Creation boundary: safely replace an existing Odysseus-owned compatibility shim."""
+    shim_file = tmp_path / "bin/llama-server"
+    shim_file.parent.mkdir(parents=True, exist_ok=True)
+    shim_file.write_text(f"#!/bin/bash\n{ODYSSEUS_LLAMA_SHIM_MARKER}\n# old shim body\n", encoding="utf-8")
+
+    lines = []
+    _append_llama_server_shim_definition_lines(lines)
+    lines.append('_odysseus_create_llama_shim')
+
+    res = _run_shell(tmp_path, lines)
+    assert res.returncode == 0
+    new_content = shim_file.read_text(encoding="utf-8")
+    assert ODYSSEUS_LLAMA_SHIM_MARKER in new_content
+    assert "exec python3 -m llama_cpp.server" in new_content
+
+
+def test_odysseus_create_llama_shim_secure_temp_prevents_symlink_attack(tmp_path):
+    """Creation boundary: secure mktemp prevents predictable $$ temp symlink attacks.
+
+    Given:
+    - A victim file at ~/victim containing DO_NOT_TOUCH
+    - A pre-created symlink at ~/bin/.llama-server.odysseus.tmp.$$ pointing to ~/victim
+    - _odysseus_create_llama_shim executed within that exact same bash PID ($$)
+
+    Prove:
+    - ~/victim remains unmodified
+    - Pre-existing crafted symlink is NOT followed or overwritten
+    - Legitimate shim is successfully created at ~/bin/llama-server via secure temp path
+    """
+    lines = [
+        'victim="$HOME/victim"',
+        'printf "DO_NOT_TOUCH\\n" > "$victim"',
+        'mkdir -p "$HOME/bin"',
+        'old_tmp="$HOME/bin/.llama-server.odysseus.tmp.$$";',
+        'ln -s "$victim" "$old_tmp"',
+    ]
+    _append_llama_server_shim_definition_lines(lines)
+    lines.append('_odysseus_create_llama_shim')
+    lines.append('res=$?')
+    lines.append('echo "CREATION_RES=$res"')
+
+    res = _run_shell(tmp_path, lines)
+    assert res.returncode == 0
+    assert "CREATION_RES=0" in res.stdout
+
+    victim_file = tmp_path / "victim"
+    assert victim_file.exists()
+    assert victim_file.read_text(encoding="utf-8") == "DO_NOT_TOUCH\n", "Victim file must remain untouched"
+
+    target_shim = tmp_path / "bin/llama-server"
+    assert target_shim.exists()
+    assert ODYSSEUS_LLAMA_SHIM_MARKER in target_shim.read_text(encoding="utf-8")
+    assert target_shim.stat().st_mode & 0o111 != 0, "Shim must be executable"
+
+
+def test_stale_odysseus_shim_removed_when_python_server_unusable(tmp_path):
+    """Scenario 1: Stale Odysseus shim + missing Python server -> owned shim removed."""
+    shim_file = tmp_path / "bin/llama-server"
+    _write_odysseus_shim(shim_file)
+
+    lines = []
+    _append_llama_server_shim_reconciliation_lines(lines)
+
+    res = _run_shell(tmp_path, lines)
+    assert res.returncode == 0
+    assert not shim_file.exists(), "Stale Odysseus shim must be removed when python server is unavailable"
+
+
+def test_healthy_odysseus_shim_retained_when_no_native_server(tmp_path):
+    """Scenario 2: Healthy Odysseus shim + no native server -> shim preserved."""
+    shim_file = tmp_path / "bin/llama-server"
+    _write_odysseus_shim(shim_file)
+    mock_py_dir = tmp_path / "mock_bin"
+    _write_mock_python3(mock_py_dir / "python3", succeeds=True)
+
+    lines = []
+    _append_llama_server_shim_reconciliation_lines(lines)
+
+    res = _run_shell(tmp_path, lines, path=f"{git_bash_path(mock_py_dir)}:{os.environ.get('PATH', '')}")
+    assert res.returncode == 0
+    assert shim_file.exists(), "Healthy Odysseus shim must be preserved when no native server is present"
+
+
+def test_reconciliation_does_not_retire_healthy_shim_for_broken_native_binary(tmp_path):
+    """Reconciliation health check: a broken native candidate executable must NOT retire
+    a healthy Odysseus Python shim."""
+    shim_file = tmp_path / "bin/llama-server"
+    _write_odysseus_shim(shim_file)
+    _write_mock_native(tmp_path / "llama.cpp/build/bin/llama-server", succeeds=False)
+    mock_py_dir = tmp_path / "mock_bin"
+    _write_mock_python3(mock_py_dir / "python3", succeeds=True)
+
+    lines = [f'export PATH="$HOME/llama.cpp/build/bin:{git_bash_path(mock_py_dir)}:$PATH"']
+    _append_llama_server_shim_reconciliation_lines(lines)
+
+    res = _run_shell(tmp_path, lines)
+    assert res.returncode == 0
+    assert shim_file.exists(), "Healthy Odysseus shim must NOT be retired when native binary is broken"
+
+
+def test_odysseus_shim_retired_when_native_server_healthy(tmp_path):
+    """Scenario 3: Odysseus shim + healthy native server available -> native server wins."""
+    shim_file = tmp_path / "bin/llama-server"
+    _write_odysseus_shim(shim_file)
+    _write_mock_native(tmp_path / "llama.cpp/build/bin/llama-server", succeeds=True)
+
+    lines = ['export PATH="$HOME/bin:$HOME/llama.cpp/build/bin:$PATH"']
+    _append_llama_server_shim_reconciliation_lines(lines)
+
+    res = _run_shell(tmp_path, lines)
+    assert res.returncode == 0
+    assert not shim_file.exists(), "Odysseus shim must be retired when healthy native server is available"
+
+
+def test_user_owned_bin_llama_server_never_touched(tmp_path):
+    """Scenario 4: Arbitrary user ~/bin/llama-server -> untouched and not deprioritized."""
+    user_bin = tmp_path / "bin/llama-server"
+    user_bin.parent.mkdir(parents=True, exist_ok=True)
+    user_bin.write_text("#!/bin/bash\n# Custom user build script\nexec /opt/custom/llama-server\n", encoding="utf-8")
+    user_bin.chmod(0o755)
+
+    _write_mock_native(tmp_path / "llama.cpp/build/bin/llama-server", succeeds=True)
+
+    lines = ['export PATH="$HOME/bin:$HOME/llama.cpp/build/bin:$PATH"']
+    _append_llama_server_shim_reconciliation_lines(lines)
+
+    res = _run_shell(tmp_path, lines)
+    assert res.returncode == 0
+    assert user_bin.exists(), "User-owned ~/bin/llama-server must never be deleted"
+    assert "Custom user build script" in user_bin.read_text(encoding="utf-8")
+
+
+def test_reconciliation_ignores_native_binary_outside_effective_path(tmp_path):
+    """Native candidate outside effective PATH must NOT be considered."""
+    shim_file = tmp_path / "bin/llama-server"
+    _write_odysseus_shim(shim_file)
+    # Native binary placed in /usr_custom/bin which is NOT in PATH
+    _write_mock_native(tmp_path / "usr_custom/bin/llama-server", succeeds=True)
+    mock_py_dir = tmp_path / "mock_bin"
+    _write_mock_python3(mock_py_dir / "python3", succeeds=True)
+
+    lines = []
+    _append_llama_server_shim_reconciliation_lines(lines)
+
+    # PATH only contains ~/bin and mock_py_dir — usr_custom is omitted
+    path = f"{git_bash_path(tmp_path / 'bin')}:{git_bash_path(mock_py_dir)}"
+    res = _run_shell(tmp_path, lines, path=path)
+    assert res.returncode == 0
+    assert shim_file.exists(), "Shim must NOT be retired by native binary outside effective PATH"
+
+
+def test_reconciliation_skips_home_bin_with_or_without_trailing_slash(tmp_path):
+    """Reconciliation must exclude $HOME/bin and $HOME/bin/ from native candidate discovery."""
+    shim_file = tmp_path / "bin/llama-server"
+    _write_odysseus_shim(shim_file)
+    mock_py_dir = tmp_path / "mock_bin"
+    _write_mock_python3(mock_py_dir / "python3", succeeds=True)
+
+    lines = []
+    _append_llama_server_shim_reconciliation_lines(lines)
+
+    # PATH has ~/bin/ with trailing slash
+    path = f"{git_bash_path(tmp_path / 'bin')}/:{git_bash_path(mock_py_dir)}"
+    res = _run_shell(tmp_path, lines, path=path)
+    assert res.returncode == 0
+    assert shim_file.exists(), "Shim must NOT rediscover itself when PATH has trailing slash in $HOME/bin/"
+
+
+def test_reconciliation_does_not_discover_cwd_executable_when_path_omits_dot(tmp_path):
+    """Reconciliation must NOT synthesize an empty/trailing PATH element that searches cwd.
+
+    Given:
+    - Current working directory contains an executable `llama-server`
+    - Healthy Odysseus Python shim in ~/bin/llama-server
+    - PATH does NOT contain '.' or cwd
+    - No other healthy native binary on PATH
+
+    Prove:
+    - Reconciliation does not discover ./llama-server
+    - The healthy Odysseus shim in ~/bin/llama-server is preserved
+    """
+    shim_file = tmp_path / "bin/llama-server"
+    _write_odysseus_shim(shim_file)
+    mock_py_dir = tmp_path / "mock_bin"
+    _write_mock_python3(mock_py_dir / "python3", succeeds=True)
+
+    cwd_dir = tmp_path / "workdir"
+    _write_mock_native(cwd_dir / "llama-server", succeeds=True)
+
+    lines = []
+    _append_llama_server_shim_reconciliation_lines(lines)
+
+    path = f"{git_bash_path(tmp_path / 'bin')}:{git_bash_path(mock_py_dir)}:/usr/bin:/bin"
+    res = _run_shell(tmp_path, lines, cwd=cwd_dir, path=path)
+    assert res.returncode == 0
+    assert shim_file.exists(), "Healthy Odysseus shim must NOT be retired by finding ./llama-server in cwd"
+
+
+def test_reconciliation_discovers_healthy_native_on_explicit_other_path_entry(tmp_path):
+    """Reconciliation must discover healthy native binaries placed on explicit PATH entries."""
+    shim_file = tmp_path / "bin/llama-server"
+    _write_odysseus_shim(shim_file)
+    custom_opt_bin = tmp_path / "custom_opt/bin"
+    _write_mock_native(custom_opt_bin / "llama-server", succeeds=True)
+
+    lines = []
+    _append_llama_server_shim_reconciliation_lines(lines)
+
+    path = f"{git_bash_path(tmp_path / 'bin')}:{git_bash_path(custom_opt_bin)}:/usr/bin:/bin"
+    res = _run_shell(tmp_path, lines, path=path)
+    assert res.returncode == 0
+    assert not shim_file.exists(), "Odysseus shim must be retired when healthy native server is on explicit PATH entry"
+
+
+def test_reconciliation_safely_iterates_path_with_glob_metacharacters(tmp_path):
+    """Reconciliation must not perform pathname expansion / globbing on PATH components.
+
+    Given:
+    - A directory path containing shell glob characters: ~/custom[dir]/bin
+    - A native binary ~/custom[dir]/bin/llama-server
+    - A working directory containing matching glob entries
+    - An existing Odysseus shim in ~/bin/llama-server
+
+    Prove:
+    - Reconciliation correctly parses the literal path without expanding [dir]
+    - Discovers the healthy native server and retires the Odysseus shim
+    """
+    shim_file = tmp_path / "bin/llama-server"
+    _write_odysseus_shim(shim_file)
+
+    glob_dir = tmp_path / "custom[dir]/bin"
+    _write_mock_native(glob_dir / "llama-server", succeeds=True)
+
+    cwd_dir = tmp_path / "cwd"
+    cwd_dir.mkdir(parents=True, exist_ok=True)
+    (cwd_dir / "customd").mkdir(parents=True, exist_ok=True)
+    (cwd_dir / "customi").mkdir(parents=True, exist_ok=True)
+
+    lines = []
+    _append_llama_server_shim_reconciliation_lines(lines)
+
+    path = f"{git_bash_path(tmp_path / 'bin')}:{git_bash_path(glob_dir)}:/usr/bin:/bin"
+    res = _run_shell(tmp_path, lines, cwd=cwd_dir, path=path)
+    assert res.returncode == 0
+    assert not shim_file.exists(), "Reconciliation must discover literal bracketed PATH without globbing"
+
+
+def test_fallback_install_success_creates_usable_launcher(tmp_path):
+    """Scenario 5: No native + no Python server -> fallback install succeeds
+    -> usable compatibility launcher exists and resolves command -v llama-server."""
+    mock_py_dir = tmp_path / "mock_bin"
+    _write_mock_python3(mock_py_dir / "python3", succeeds=True)
+
+    lines = [f'export PATH="$HOME/bin:{git_bash_path(mock_py_dir)}:$PATH"']
+    _append_llama_server_shim_definition_lines(lines)
+    _append_llama_server_finalization_lines(lines)
+    lines.append('command -v llama-server')
+
+    res = _run_shell(tmp_path, lines)
+    assert res.returncode == 0
+    created_shim = tmp_path / "bin/llama-server"
+    assert created_shim.exists()
+    assert ODYSSEUS_LLAMA_SHIM_MARKER in created_shim.read_text(encoding="utf-8")
+    assert str(tmp_path / "bin/llama-server") in res.stdout
+
+
+def test_failed_fallback_sets_preflight_exit_127(tmp_path):
+    """Scenario 6: No native + failed fallback installation -> ODYSSEUS_PREFLIGHT_EXIT=127."""
+    lines = [
+        'ODYSSEUS_PREFLIGHT_EXIT=""',
+        'export PATH="$HOME/bin:$PATH"',
+    ]
+    _append_llama_server_shim_definition_lines(lines)
+    _append_llama_server_finalization_lines(lines)
+    lines.append('echo "EXIT=$ODYSSEUS_PREFLIGHT_EXIT"')
+
+    res = _run_shell(tmp_path, lines)
+    assert res.returncode == 0
+    assert "EXIT=127" in res.stdout
+
+
+def test_python_server_readiness_requires_llama_cpp_server_module():
+    """Scenario 7 & 8: Python server readiness requires `import llama_cpp.server`,
+    while GPU offload capability checks still use base `import llama_cpp`."""
+    routes_source = _read("routes/cookbook_routes.py")
+    helpers_source = _read("routes/cookbook_helpers.py")
+
+    # Windows fallback readiness check requires .server
+    assert 'try { python -c "import llama_cpp.server" 2>$null } catch {}' in routes_source
+    # Termux fallback readiness check requires .server
+    assert 'if ! python3 -c "import llama_cpp.server" 2>/dev/null; then' in routes_source
+    # Build-failure fallback pip check requires .server
+    assert 'if [ -z "$_odysseus_llama_bin" ] && ! python3 -c "import llama_cpp.server" 2>/dev/null; then' in routes_source
+    # Shim reconciliation and post-install finalization require .server
+    assert 'elif python3 -c "import llama_cpp.server" 2>/dev/null; then' in helpers_source
+    assert 'if [ -z "$_odysseus_llama_bin" ] && python3 -c "import llama_cpp.server" 2>/dev/null; then' in helpers_source
+
+    # GPU offload support checks still intentionally inspect base llama_cpp
+    assert 'llama_cpp.llama_supports_gpu_offload()' in routes_source
+
+
+def test_darwin_native_build_does_not_force_write_bin_symlink():
+    """Darwin source build must not overwrite ~/bin/llama-server with a forced symlink."""
+    source = _read("routes/cookbook_routes.py")
+
+    assert "&& ln -sf ~/llama.cpp/build/bin/llama-server ~/bin/llama-server" not in source
+    assert 'cmake --build build -j"$NPROC" --target llama-server' in source
+
+
+@pytest.mark.asyncio
+async def test_generated_runner_preflight_retires_stale_shim_before_backend_selection(monkeypatch, tmp_path):
+    """Integration regression: The actual generated Cookbook runner must execute
+    ownership reconciliation BEFORE evaluating the backend selection (`if Termux ... elif ! command -v llama-server`).
+
+    Given:
+    - Executable stale Odysseus shim in ~/bin/llama-server (with canonical marker)
+    - Healthy native llama-server in ~/llama.cpp/build/bin/llama-server
+    - ~/bin before ~/llama.cpp/build/bin on PATH
+
+    Prove:
+    1. Generated runner contains reconciliation shell lines strictly before the backend branching
+    2. Executing the generated runner script retires the stale ~/bin/llama-server shim
+    3. Resolves to the healthy native binary without attempting source build or pip fallback.
+    """
+    from unittest.mock import AsyncMock, MagicMock
+    from routes.cookbook_helpers import ServeRequest
+    import routes.cookbook_routes as cookbook_routes
+    from starlette.requests import Request
+
+    fake_home = tmp_path / "fake_home"
+    stale_shim = fake_home / "bin/llama-server"
+    _write_odysseus_shim(stale_shim)
+    native_bin = fake_home / "llama.cpp/build/bin/llama-server"
+    _write_mock_native(native_bin, succeeds=True)
+
+    # Mock environment to generate the real runner file with isolated DB
+    tmux_log_dir = tmp_path / "tmux_logs"
+    tmux_log_dir.mkdir(parents=True)
+    monkeypatch.setattr(cookbook_routes, "TMUX_LOG_DIR", tmux_log_dir)
+    monkeypatch.setattr(cookbook_routes, "require_admin", lambda request: None)
+    monkeypatch.setattr(cookbook_routes, "_binary_available", AsyncMock(return_value=True))
+    monkeypatch.setattr(cookbook_routes, "load_stored_hf_token", lambda **kwargs: "")
+    monkeypatch.setattr("core.database.SessionLocal", MagicMock())
+
+    class _Proc:
+        returncode = 0
+        stderr = None
+        async def wait(self):
+            return 0
+    monkeypatch.setattr(cookbook_routes.asyncio, "create_subprocess_shell", AsyncMock(return_value=_Proc()))
+
+    router = cookbook_routes.setup_cookbook_routes()
+    endpoint = None
+    for route in router.routes:
+        if route.path == "/api/model/serve" and "POST" in route.methods:
+            endpoint = route.endpoint
+            break
+    assert endpoint is not None
+
+    from types import SimpleNamespace
+    http_req = Request({
+        "type": "http",
+        "method": "POST",
+        "path": "/api/model/serve",
+        "headers": [],
+        "app": SimpleNamespace(state=SimpleNamespace(auth_manager=SimpleNamespace(
+            is_configured=True, is_admin=lambda user: user == "admin"))),
+    })
+    http_req.state.current_user = "admin"
+    http_req.state.local_model_authority = True
+
+    req = ServeRequest(
+        repo_id="test/repo",
+        cmd="llama-server -m /dummy/model.gguf --port 8085",
+        remote_host=None,
+    )
+
+    resp = await endpoint(http_req, req)
+    assert resp["ok"] is True
+    session_id = resp["session_id"]
+    runner_file = tmux_log_dir / f"{session_id}_run.sh"
+    assert runner_file.exists()
+    runner_script = runner_file.read_text(encoding="utf-8")
+
+    # Assert structural ordering in the generated script:
+    # Reconciliation must appear strictly BEFORE the Termux/backend check
+    recon_idx = runner_script.find("Native llama-server detected — retiring Odysseus compatibility shim in ~/bin")
+    termux_idx = runner_script.find("if [ -d /data/data/com.termux ]; then")
+    backend_idx = runner_script.find('elif [ -z "$_odysseus_llama_bin" ]; then')
+    assert recon_idx != -1, "Reconciliation block missing from generated runner"
+    assert termux_idx != -1, "Termux check missing from generated runner"
+    assert backend_idx != -1, "Backend selection missing from generated runner"
+    assert recon_idx < termux_idx < backend_idx, "Reconciliation must execute strictly before backend selection"
+
+    # Execute the generated runner preflight in a real subshell
+    res = _run_shell(fake_home, [str(runner_file)], path=f"{git_bash_path(fake_home / 'bin')}:{git_bash_path(fake_home / 'llama.cpp/build/bin')}:/usr/bin:/bin")
+    assert res.returncode == 0
+    assert not stale_shim.exists(), "Stale ~/bin/llama-server must have been retired by the generated preflight"
+    assert "RUN_NATIVE: -m /dummy/model.gguf --port 8085" in res.stdout
+    assert "Native llama-server not found — building from source" not in res.stdout
+
+
+def test_broken_first_native_candidate_healthy_second_candidate_executes_healthy(tmp_path):
+    """Regression finding P2: When PATH contains an earlier broken candidate followed by a
+    healthy candidate, reconciliation and launch must select and execute the healthy candidate."""
+    from core.platform_compat import git_bash_path
+    broken_dir = tmp_path / "broken_bin"
+    _write_mock_native(broken_dir / "llama-server", succeeds=False)
+
+    healthy_dir = tmp_path / "healthy_bin"
+    _write_mock_native(healthy_dir / "llama-server", succeeds=True)
+
+    shim_file = tmp_path / "bin/llama-server"
+    _write_odysseus_shim(shim_file)
+
+    lines = [
+        f'export PATH="$HOME/bin:{git_bash_path(broken_dir)}:{git_bash_path(healthy_dir)}:/usr/bin:/bin"',
+    ]
+    _append_llama_server_shim_reconciliation_lines(lines)
+    lines.append('echo "RESOLVED=$_odysseus_llama_bin"')
+    lines.append('command -v llama-server')
+    lines.append('llama-server -m /dummy/model.gguf --port 8080')
+
+    res = _run_shell(tmp_path, lines)
+    assert res.returncode == 0
+    assert not shim_file.exists(), "Stale Odysseus shim must be retired"
+    assert f"RESOLVED={git_bash_path(healthy_dir / 'llama-server')}" in res.stdout
+    assert str(healthy_dir / "llama-server") in res.stdout
+    assert "RUN_NATIVE: -m /dummy/model.gguf --port 8080" in res.stdout
+    assert "error: missing libllama.so" not in res.stdout
+
+
+def test_paths_with_spaces_broken_first_healthy_second_launch(tmp_path):
+    """Regression: Paths containing spaces must be safely resolved, prioritized and launched."""
+    from core.platform_compat import git_bash_path
+    broken_dir = tmp_path / "broken path with spaces/bin"
+    _write_mock_native(broken_dir / "llama-server", succeeds=False)
+
+    healthy_dir = tmp_path / "healthy path with spaces/bin"
+    _write_mock_native(healthy_dir / "llama-server", succeeds=True)
+
+    shim_file = tmp_path / "bin/llama-server"
+    _write_odysseus_shim(shim_file)
+
+    model_dir = tmp_path / "models with spaces"
+    model_dir.mkdir(parents=True, exist_ok=True)
+    model_file = model_dir / "my model.gguf"
+    model_file.write_text("dummy", encoding="utf-8")
+
+    lines = [
+        f'export PATH="$HOME/bin:{git_bash_path(broken_dir)}:{git_bash_path(healthy_dir)}:/usr/bin:/bin"',
+    ]
+    _append_llama_server_shim_reconciliation_lines(lines)
+    lines.append(f'llama-server -m "{git_bash_path(model_file)}" --port 8080')
+
+    res = _run_shell(tmp_path, lines)
+    assert res.returncode == 0
+    assert not shim_file.exists(), "Stale shim must be retired"
+    assert f"RUN_NATIVE: -m {git_bash_path(model_file)} --port 8080" in res.stdout
+    assert "error: missing libllama.so" not in res.stdout
+
+
+def test_command_hash_invalidation_switches_from_cached_broken_to_healthy(tmp_path):
+    """Regression: Command cache invalidation (`hash -r`) must prevent bash from reusing a
+    previously cached lookup to an earlier broken executable."""
+    from core.platform_compat import git_bash_path
+    broken_dir = tmp_path / "broken_bin"
+    _write_mock_native(broken_dir / "llama-server", succeeds=False)
+
+    healthy_dir = tmp_path / "healthy_bin"
+    _write_mock_native(healthy_dir / "llama-server", succeeds=True)
+
+    lines = [
+        # Populate bash hash cache with broken binary
+        f'export PATH="{git_bash_path(broken_dir)}:/usr/bin:/bin"',
+        'llama-server >/dev/null 2>&1 || true',
+        'hash',
+        # Now add healthy candidate later in PATH and run reconciliation
+        f'export PATH="$HOME/bin:{git_bash_path(broken_dir)}:{git_bash_path(healthy_dir)}:/usr/bin:/bin"',
+    ]
+    _append_llama_server_shim_reconciliation_lines(lines)
+    lines.append('llama-server -m /dummy/test.gguf')
+    lines.append('hash')
+
+    res = _run_shell(tmp_path, lines)
+    assert res.returncode == 0
+    assert "RUN_NATIVE: -m /dummy/test.gguf" in res.stdout
+    # Verify hash table ended with healthy binary, not broken
+    assert str(healthy_dir / "llama-server") in res.stdout
+
+
+def test_broken_native_candidate_falls_back_to_healthy_python_compatibility_shim(tmp_path):
+    """Regression: A broken native candidate must not shadow or retire a healthy Python shim."""
+    from core.platform_compat import git_bash_path
+    broken_dir = tmp_path / "broken_bin"
+    _write_mock_native(broken_dir / "llama-server", succeeds=False)
+
+    shim_file = tmp_path / "bin/llama-server"
+    _write_odysseus_shim(shim_file)
+
+    mock_py_dir = tmp_path / "mock_bin"
+    _write_mock_python3(mock_py_dir / "python3", succeeds=True)
+
+    lines = [
+        f'export PATH="$HOME/bin:{git_bash_path(broken_dir)}:{git_bash_path(mock_py_dir)}:/usr/bin:/bin"',
+    ]
+    _append_llama_server_shim_reconciliation_lines(lines)
+    lines.append('echo "RESOLVED=$_odysseus_llama_bin"')
+    lines.append('command -v llama-server')
+
+    res = _run_shell(tmp_path, lines)
+    assert res.returncode == 0
+    assert shim_file.exists(), "Healthy Python compatibility shim must be preserved"
+    assert f"RESOLVED={git_bash_path(shim_file)}" in res.stdout
+    assert str(shim_file) in res.stdout
+
+
+def test_broken_native_and_missing_python_backend_sets_preflight_127(tmp_path):
+    """Regression: When the only native candidates are broken and Python server is unavailable,
+    finalization must fail closed with ODYSSEUS_PREFLIGHT_EXIT=127 rather than claiming readiness."""
+    from core.platform_compat import git_bash_path
+    broken_dir = tmp_path / "broken_bin"
+    _write_mock_native(broken_dir / "llama-server", succeeds=False)
+
+    mock_py_dir = tmp_path / "mock_bin"
+    _write_mock_python3(mock_py_dir / "python3", succeeds=False)
+
+    lines = [
+        'ODYSSEUS_PREFLIGHT_EXIT=""',
+        f'export PATH="$HOME/bin:{git_bash_path(broken_dir)}:{git_bash_path(mock_py_dir)}:/usr/bin:/bin"',
+    ]
+    _append_llama_server_shim_definition_lines(lines)
+    _append_llama_server_shim_reconciliation_lines(lines)
+    _append_llama_server_finalization_lines(lines)
+    lines.append('echo "EXIT=$ODYSSEUS_PREFLIGHT_EXIT"')
+
+    res = _run_shell(tmp_path, lines)
+    assert res.returncode == 0
+    assert "EXIT=127" in res.stdout
+
+
+@pytest.mark.asyncio
+async def test_generated_runner_broken_first_healthy_second_e2e(monkeypatch, tmp_path):
+    """Integration regression: The full generated Cookbook runner script must launch the healthy
+    native candidate even when an earlier broken candidate exists on PATH."""
+    from unittest.mock import AsyncMock, MagicMock
+    from types import SimpleNamespace
+    from routes.cookbook_helpers import ServeRequest
+    import routes.cookbook_routes as cookbook_routes
+    from starlette.requests import Request
+    from core.platform_compat import git_bash_path
+
+    fake_home = tmp_path / "fake_home"
+    stale_shim = fake_home / "bin/llama-server"
+    _write_odysseus_shim(stale_shim)
+
+    broken_bin = fake_home / "broken/bin/llama-server"
+    _write_mock_native(broken_bin, succeeds=False)
+
+    healthy_bin = fake_home / "healthy/bin/llama-server"
+    _write_mock_native(healthy_bin, succeeds=True)
+
+    tmux_log_dir = tmp_path / "tmux_logs"
+    tmux_log_dir.mkdir(parents=True)
+    monkeypatch.setattr(cookbook_routes, "TMUX_LOG_DIR", tmux_log_dir)
+    monkeypatch.setattr(cookbook_routes, "require_admin", lambda request: None)
+    monkeypatch.setattr(cookbook_routes, "_binary_available", AsyncMock(return_value=True))
+    monkeypatch.setattr(cookbook_routes, "load_stored_hf_token", lambda **kwargs: "")
+    monkeypatch.setattr("core.database.SessionLocal", MagicMock())
+
+    class _Proc:
+        returncode = 0
+        stderr = None
+        async def wait(self):
+            return 0
+    monkeypatch.setattr(cookbook_routes.asyncio, "create_subprocess_shell", AsyncMock(return_value=_Proc()))
+
+    router = cookbook_routes.setup_cookbook_routes()
+    endpoint = None
+    for route in router.routes:
+        if route.path == "/api/model/serve" and "POST" in route.methods:
+            endpoint = route.endpoint
+            break
+    assert endpoint is not None
+
+    http_req = Request({
+        "type": "http",
+        "method": "POST",
+        "path": "/api/model/serve",
+        "headers": [],
+        "app": SimpleNamespace(state=SimpleNamespace(auth_manager=SimpleNamespace(
+            is_configured=True, is_admin=lambda user: user == "admin"))),
+    })
+    http_req.state.current_user = "admin"
+    http_req.state.local_model_authority = True
+
+    req = ServeRequest(
+        repo_id="test/repo",
+        cmd="llama-server -m /dummy/model.gguf --port 8090",
+        remote_host=None,
+    )
+
+    resp = await endpoint(http_req, req)
+    assert resp["ok"] is True
+    session_id = resp["session_id"]
+    runner_file = tmux_log_dir / f"{session_id}_run.sh"
+    assert runner_file.exists()
+
+    # Broken is earlier on PATH than healthy
+    res = _run_shell(
+        fake_home,
+        [str(runner_file)],
+        path=f"{git_bash_path(fake_home / 'bin')}:{git_bash_path(broken_bin.parent)}:{git_bash_path(healthy_bin.parent)}:/usr/bin:/bin",
+    )
+    assert res.returncode == 0
+    assert not stale_shim.exists(), "Stale shim must have been retired"
+    assert "RUN_NATIVE: -m /dummy/model.gguf --port 8090" in res.stdout
+    assert "error: missing libllama.so" not in res.stdout
+    assert "Native llama-server not found — building from source" not in res.stdout
diff --git a/tests/test_database_test_isolation.py b/tests/test_database_test_isolation.py
index f0e72e27a..308ff18ec 100644
--- a/tests/test_database_test_isolation.py
+++ b/tests/test_database_test_isolation.py
@@ -85,6 +85,9 @@ def test_actual_tests_restore_process_state_and_ignore_inherited_database(tmp_pa
         import os
         import sys
         import pytest
+        import tempfile
+        from pathlib import Path
+        from sqlalchemy.engine import make_url
 
         def snapshot():
             import core
@@ -107,7 +110,11 @@ def test_actual_tests_restore_process_state_and_ignore_inherited_database(tmp_pa
         class StateGuard:
             def pytest_sessionstart(self):
                 self.initial = snapshot()
-                assert self.initial[0] == "sqlite:///:memory:"
+                assert self.initial[0] != inherited_url
+                self.default_path = Path(make_url(self.initial[0]).database)
+                assert self.default_path.is_file()
+                assert self.default_path.name == "collection.db"
+                assert self.default_path.parent == Path(tempfile.gettempdir()).parent
 
             def pytest_collection_finish(self):
                 assert snapshot() == self.initial, "collection changed database bindings"
@@ -118,9 +125,11 @@ def test_actual_tests_restore_process_state_and_ignore_inherited_database(tmp_pa
                 assert snapshot() == self.initial, "test leaked database or module state"
 
         inherited_url = os.environ["DATABASE_URL"]
+        guard = StateGuard()
         result = pytest.main(["-q", "-p", "no:cacheprovider", *sys.argv[1:]],
-                             plugins=[StateGuard()])
+                             plugins=[guard])
         assert os.environ["DATABASE_URL"] == inherited_url
+        assert not guard.default_path.exists(), "collection database outlived its runtime"
         raise SystemExit(result)
     ''')
     result = subprocess.run(
diff --git a/tests/test_delete_user_invalidates_token_cache.py b/tests/test_delete_user_invalidates_token_cache.py
index 91be50e93..2a534f89e 100644
--- a/tests/test_delete_user_invalidates_token_cache.py
+++ b/tests/test_delete_user_invalidates_token_cache.py
@@ -11,6 +11,8 @@ cache from the DB, where the rows are already gone, and the token is rejected.
 """
 import asyncio
 import types
+import pytest
+from fastapi import HTTPException
 
 from routes.auth_routes import setup_auth_routes, DeleteUserRequest
 
@@ -60,12 +62,9 @@ def test_refused_delete_does_not_invalidate_cache():
     invalidations = []
     router = setup_auth_routes(_auth_manager(delete_result=False))
     handler = _handler(router)
-    try:
+    with pytest.raises(HTTPException) as refused:
         asyncio.run(handler(DeleteUserRequest(username="admin"), _fake_request(invalidations)))
-        raised = False
-    except Exception:
-        raised = True
-    assert raised, "a refused delete should raise (HTTP 400)"
+    assert refused.value.status_code == 400
     assert invalidations == [], "a refused delete must not touch the token cache"
 
 
diff --git a/tests/test_email_list_stale_while_revalidate.py b/tests/test_email_list_stale_while_revalidate.py
index 4fe22a75d..13f347f5a 100644
--- a/tests/test_email_list_stale_while_revalidate.py
+++ b/tests/test_email_list_stale_while_revalidate.py
@@ -56,14 +56,35 @@ def test_cold_start_with_no_snapshot_and_no_index_is_not_marked_stale(monkeypatc
     # falls through to the real (blocking) path, so the response is not a
     # stale-marked snapshot.
     endpoint = _list_endpoint()
+    calls = []
+
+    class EmptyMailbox:
+        def select(self, folder, readonly=False):
+            calls.append(("select", folder, readonly))
+            return "OK", [b"0"]
+
+        def uid(self, command, *args):
+            assert command == "SEARCH"
+            calls.append(("search", args))
+            return "OK", [b""]
+
+        def logout(self):
+            pass
+
+    def connect(*args, **kwargs):
+        calls.append(("connect", args, kwargs))
+        return EmptyMailbox()
+
+    monkeypatch.setattr(er, "_start_poller", lambda: None)
     monkeypatch.setattr(er, "_email_index_list", lambda *a, **k: ([], 0, None))
-    monkeypatch.setattr(er, "_imap_connect", lambda *a, **k: (_ for _ in ()).throw(
-        RuntimeError("offline test")))
-    try:
-        result = _call(endpoint)
-    except Exception:
-        return  # the cold path legitimately attempted real IMAP
+    monkeypatch.setattr(er, "_imap_connect", connect)
+    result = _call(endpoint)
+    assert [call[0] for call in calls] == ["connect", "select", "search"]
+    assert calls[1][2] is True
     assert isinstance(result, dict)
+    assert result["emails"] == []
+    assert result["total"] == 0
+    assert not result.get("error")
     assert not (result.get("sync") or {}).get("stale")
 
 
diff --git a/tests/test_mcp_manager.py b/tests/test_mcp_manager.py
index a879f95eb..11e655422 100644
--- a/tests/test_mcp_manager.py
+++ b/tests/test_mcp_manager.py
@@ -39,3 +39,74 @@ def test_http_transport_routes_to_start_http_connect():
         result = asyncio.run(mgr.connect_server("id1", "n", "http", url="https://x/mcp"))
     assert result == "ROUTED"
     m.assert_called_once()
+
+
+def _patch_http_transport(monkeypatch, initialize):
+    """Fake the Streamable HTTP transport and record, for each close, whether
+    it ran in the task that opened it (AnyIO requires that)."""
+    import contextlib
+    import mcp
+    import mcp.client.streamable_http as streamable_http
+    import src.mcp_oauth as mcp_oauth
+
+    closes = []
+
+    @contextlib.asynccontextmanager
+    async def fake_client(url, auth=None):
+        opener = asyncio.current_task()
+        try:
+            yield (object(), object(), lambda: None)
+        finally:
+            closes.append(asyncio.current_task() is opener)
+
+    class FakeSession:
+        def __init__(self, *a):
+            pass
+
+        async def __aenter__(self):
+            return self
+
+        async def __aexit__(self, *exc):
+            return False
+
+    FakeSession.initialize = initialize
+    monkeypatch.setattr(streamable_http, "streamablehttp_client", fake_client)
+    monkeypatch.setattr(mcp, "ClientSession", FakeSession)
+    monkeypatch.setattr(mcp_oauth, "build_provider", lambda *a, **k: None)
+    return closes
+
+
+def test_failed_http_connect_closes_transport_in_its_own_task(monkeypatch):
+    """#5518: a failed connect left its transport for the GC, which exited the
+    AnyIO cancel scope from the wrong task and kept the event loop spinning."""
+    async def initialize(self):
+        raise RuntimeError("Session terminated")
+
+    closes = _patch_http_transport(monkeypatch, initialize)
+    mgr = McpManager()
+
+    ok = asyncio.run(mgr._start_http_connect("s1", "remote", "http://x/mcp/", wait=2))
+
+    assert ok is False
+    assert mgr._connections["s1"]["status"] == "error"
+    assert "Session terminated" in mgr._connections["s1"]["error"]
+    assert closes == [True]
+    assert "s1" not in mgr._stacks
+
+
+def test_cancelled_http_connect_closes_transport(monkeypatch):
+    """Deleting a server mid-connect cancels the connect task; the transport
+    must still be closed rather than leaked."""
+    async def initialize(self):
+        await asyncio.sleep(3600)
+
+    closes = _patch_http_transport(monkeypatch, initialize)
+    mgr = McpManager()
+
+    async def run():
+        await mgr._start_http_connect("s1", "remote", "http://x/mcp/", wait=0.05)
+        await mgr.disconnect_server("s1")
+        await asyncio.sleep(0)
+
+    asyncio.run(run())
+    assert closes == [True]
diff --git a/tests/test_session_actions_cleanup.py b/tests/test_session_actions_cleanup.py
index 221713d33..be89a29de 100644
--- a/tests/test_session_actions_cleanup.py
+++ b/tests/test_session_actions_cleanup.py
@@ -128,7 +128,7 @@ def test_auto_sort_keeps_fresh_session_while_first_response_is_pending(monkeypat
         db.close()
 
 
-def test_auto_sort_still_deletes_old_throwaway_sessions(monkeypatch):
+def test_auto_sort_keeps_old_nonempty_sessions_even_with_throwaway_title(monkeypatch):
     session_factory = _make_session_factory()
     _install_session_factory(monkeypatch, session_factory)
 
@@ -160,7 +160,8 @@ def test_auto_sort_still_deletes_old_throwaway_sessions(monkeypatch):
 
     db = session_factory()
     try:
-        assert db.query(DbSession).filter(DbSession.id == sid).first() is None
-        assert "Cleaned 1 sessions" in result
+        assert db.query(DbSession).filter(DbSession.id == sid).first() is not None
+        assert db.query(DbMessage).filter(DbMessage.session_id == sid).count() == 1
+        assert "Cleaned 0 sessions" in result
     finally:
         db.close()
diff --git a/tests/test_session_list_owner_scope.py b/tests/test_session_list_owner_scope.py
index fc973bd53..aacde8d29 100644
--- a/tests/test_session_list_owner_scope.py
+++ b/tests/test_session_list_owner_scope.py
@@ -19,6 +19,7 @@ from core.database import ChatMessage as DbMessage
 from core.database import Session as DbSession
 
 _TMPDB = tempfile.NamedTemporaryFile(suffix=".db", delete=False)
+_TMPDB.close()  # SQLite owns its connections; release the Windows file handle.
 _ENGINE = create_engine(
     f"sqlite:///{_TMPDB.name}",
     connect_args={"check_same_thread": False},
@@ -230,12 +231,14 @@ def test_patch_session_updates_persisted_cwd(monkeypatch):
 
 def test_auto_sort_skip_llm_cleans_owner_stamped_sessions_when_auth_disabled(monkeypatch):
     import routes.session_routes as sr
+    from fastapi import APIRouter
     from unittest.mock import MagicMock
 
     _stub_multipart_if_missing(monkeypatch)
     monkeypatch.setenv("AUTH_ENABLED", "false")
     monkeypatch.setattr(sr, "SessionLocal", _TS)
     monkeypatch.setattr(sr, "effective_user", lambda request: None)
+    monkeypatch.setattr(sr, "router", APIRouter(prefix="/api"))
 
     sid = str(uuid.uuid4())
     old_time = cdb.utcnow_naive() - timedelta(hours=2)
@@ -250,19 +253,12 @@ def test_auto_sort_skip_llm_cleans_owner_stamped_sessions_when_auth_disabled(mon
             endpoint_url="http://localhost",
             model="gpt-4",
             archived=False,
-            message_count=1,
+            message_count=0,
             created_at=old_time,
             updated_at=old_time,
             last_message_at=old_time,
             last_accessed=old_time,
         ))
-        db.add(DbMessage(
-            id="m-" + uuid.uuid4().hex,
-            session_id=sid,
-            role="user",
-            content="hi",
-            timestamp=old_time,
-        ))
         db.commit()
     finally:
         db.close()
@@ -270,6 +266,13 @@ def test_auto_sort_skip_llm_cleans_owner_stamped_sessions_when_auth_disabled(mon
     session = MagicMock(id=sid, name="New chat", model="gpt-4", endpoint_url="http://localhost", rag=False, archived=False)
     sm = MagicMock()
     sm.get_sessions_for_user.return_value = {sid: session}
+
+    def delete_session(session_id):
+        with _TS() as db:
+            db.query(DbSession).filter(DbSession.id == session_id).delete()
+            db.commit()
+
+    sm.delete_session.side_effect = delete_session
     router = sr.setup_session_routes(sm, {})
     endpoint = next(r.endpoint for r in router.routes
                     if getattr(r, "path", "") == "/api/sessions/auto-sort"
@@ -277,7 +280,8 @@ def test_auto_sort_skip_llm_cleans_owner_stamped_sessions_when_auth_disabled(mon
 
     result = endpoint(request=MagicMock(), skip_llm=True)
 
-    assert result["deleted_throwaway"] == 1
+    assert result["deleted_empty"] == 1
+    assert result["deleted_throwaway"] == 0
     db = _TS()
     try:
         assert db.query(DbSession).filter(DbSession.id == sid).first() is None
diff --git a/tests/test_worker_runtime.py b/tests/test_worker_runtime.py
index a91157902..b99d58817 100644
--- a/tests/test_worker_runtime.py
+++ b/tests/test_worker_runtime.py
@@ -79,3 +79,32 @@ def test_tmp_path_under_private_runtime_fits_unix_socket(tmp_path):
     # witness. A nested pytest-of-<user> basetemp made it 110 bytes under xdist.
     with socket.socket(socket.AF_UNIX) as sock:
         sock.bind(str(tmp_path / "tmux.sock"))
+
+
+def test_collection_database_survives_thread_connection_churn():
+    """TestClient threads must not lose the collection database or its schema.
+
+    SingletonThreadPool's default five-connection limit used to close the main
+    thread's in-memory database after repeated route clients. Each new thread
+    also saw an empty database. Exercise more threads than that limit and then
+    verify both compatibility factories still see all registered tables.
+    """
+    from concurrent.futures import ThreadPoolExecutor
+    from sqlalchemy import inspect
+    import core.database as database
+    import src.database as compatibility
+
+    expected = set(database.Base.metadata.tables)
+
+    def table_names(factory):
+        with factory() as session:
+            return set(inspect(session.connection()).get_table_names())
+
+    assert {"documents", "model_endpoints", "mcp_servers", "calendars"} <= expected
+    assert expected <= table_names(database.SessionLocal)
+    for _ in range(8):
+        # A fresh executor creates a fresh thread rather than reusing one.
+        with ThreadPoolExecutor(max_workers=1) as executor:
+            assert expected <= executor.submit(table_names, database.SessionLocal).result()
+    assert expected <= table_names(database.SessionLocal)
+    assert expected <= table_names(compatibility.SessionLocal)