deploy: sync backend (real general answers, casual short-circuit, classifier, chat resilience)

backend/app/services/learn_lesson_builder.py

@@ -36,10 +36,11 @@ from app.core.config import BACKEND_DIR, PROJECT_ROOT, get_settings
 
 logger = logging.getLogger(__name__)
 
-# Writable cache for lesson manifests + audio. On Hugging Face Docker the
-# backend tree is /app, so PROJECT_ROOT/public is outside the container and
-# not writable (PermissionError → opaque 500). Cache under BACKEND_DIR.
-# Tests may monkeypatch this path.
+# Writable cache for lesson manifests + audio.
+# NEVER use PROJECT_ROOT/"public" — on HF Docker WORKDIR=/app, that resolves to
+# /public which is not creatable (Permission denied → 500).
+# Prefer: LEARN_LESSON_CACHE_DIR → /app/generated/... → backend/generated/...
+# Tests may monkeypatch PUBLIC_ROOT.
 PUBLIC_ROOT = BACKEND_DIR / "generated" / "learn-anything"
 
 GROQ_URL = "https://api.groq.com/openai/v1/chat/completions"
@@ -761,25 +762,49 @@ def _lesson_lock_for(key: str) -> threading.Lock:
 
 
 def ensure_public_root() -> Path:
-    """Ensure the lesson cache directory exists and is writable."""
+    """Ensure the lesson cache directory exists and is writable on HF + local."""
+    candidates: list[Path] = []
     env = _env("LEARN_LESSON_CACHE_DIR")
-    root = Path(env) if env else PUBLIC_ROOT
-    try:
-        root.mkdir(parents=True, exist_ok=True)
-        probe = root / ".write_probe"
-        probe.write_text("ok", encoding="utf-8")
-        probe.unlink(missing_ok=True)
-        return root
-    except OSError:
-        fallback = Path(os.getenv("TMPDIR") or os.getenv("TEMP") or "/tmp") / "docdoe-learn-lessons"
+    if env:
+        candidates.append(Path(env))
+    # HF Docker WORKDIR is /app (backend tree).
+    candidates.append(Path("/app/generated/learn-anything"))
+    # Monorepo / local backend package root.
+    candidates.append(BACKEND_DIR / "generated" / "learn-anything")
+    # Tests monkeypatch PUBLIC_ROOT — honour it if already set to a temp path.
+    if PUBLIC_ROOT not in candidates:
+        candidates.insert(0, PUBLIC_ROOT)
+    # Last-resort temp (always writable for reading-mode fallbacks).
+    candidates.append(
+        Path(os.getenv("TMPDIR") or os.getenv("TEMP") or "/tmp")
+        / "docdoe-learn-lessons"
+    )
+
+    errors: list[str] = []
+    for root in candidates:
+        # Never attempt to create absolute /public (HF permission bomb).
         try:
-            fallback.mkdir(parents=True, exist_ok=True)
-            logger.warning("Lesson cache fell back to %s", fallback)
-            return fallback
+            resolved = root.resolve()
+        except OSError:
+            resolved = root
+        if str(resolved).startswith("/public") or str(resolved) == "/public":
+            errors.append(f"skip forbidden path {root}")
+            continue
+        try:
+            root.mkdir(parents=True, exist_ok=True)
+            probe = root / ".write_probe"
+            probe.write_text("ok", encoding="utf-8")
+            probe.unlink(missing_ok=True)
+            if root != PUBLIC_ROOT:
+                logger.info("Lesson cache using %s", root)
+            return root
         except OSError as exc:
-            raise LessonBuildError(
-                f"Lesson cache directory is not writable ({root}): {exc}"
-            ) from exc
+            errors.append(f"{root}: {exc}")
+            continue
+
+    raise LessonBuildError(
+        "Lesson cache directory is not writable. Tried: " + "; ".join(errors)
+    )
 
 
 def build_lesson(
@@ -835,9 +860,26 @@ def build_lesson(
         try:
             out_dir.mkdir(parents=True, exist_ok=True)
         except OSError as exc:
-            raise LessonBuildError(
-                f"Could not create lesson cache folder ({out_dir}): {exc}"
-            ) from exc
+            # Hard fallback: rebuild under /tmp so students still get a class.
+            emergency = (
+                Path(os.getenv("TMPDIR") or os.getenv("TEMP") or "/tmp")
+                / "docdoe-learn-lessons"
+                / key
+            )
+            try:
+                emergency.mkdir(parents=True, exist_ok=True)
+                out_dir = emergency
+                manifest_path = out_dir / "lesson.json"
+                logger.warning(
+                    "Lesson cache mkdir failed (%s); using emergency path %s",
+                    exc,
+                    out_dir,
+                )
+            except OSError as exc2:
+                raise LessonBuildError(
+                    f"Could not create lesson cache folder ({out_dir}): {exc}; "
+                    f"emergency also failed: {exc2}"
+                ) from exc2
         delivery_mode = "reading" if mock_mode else "audio"
         delivery_notice = (
             "Local preview: this complete starter class is available in reading mode; no generated voice is being presented as real audio."