ContentClassifier: `MAX_BUFFER_SIZE` is declared but never enforced

backend/utils/content_classifier_utils.py:20 advertises a 1 MB buffer cap as part of the "DoS protection" the class docstring touts:

MAX_BUFFER_SIZE = 1024 * 1024  # 1MB
MAX_TAG_LENGTH = 256           # Single tag max length
MAX_PATH_LENGTH = 512          # File path max length
MAX_TAG_COUNT = 100            # Max tags before stopping

A search through the same file shows the only places that mention these constants:

20:    MAX_BUFFER_SIZE = 1024 * 1024  # 1MB
23:    MAX_TAG_COUNT = 100            # Max tags before stopping
73:        if self.tag_count >= self.MAX_TAG_COUNT:

MAX_TAG_LENGTH and MAX_PATH_LENGTH are referenced (lines 63, 146); MAX_BUFFER_SIZE is never read. The classify() method (line 39) appends every incoming chunk to self.buffer with no upper bound:

def classify(self, chunk: str) -> List[Dict[str, Any]]:
    results = []
    self.buffer += chunk
    while self.buffer:
        ...

A streaming LLM that emits a very long run of "<" characters (or simply a very long single line with no recognised tag start) will grow self.buffer unboundedly while _process_non_tag_content emits 64-byte chunks back. An adversarial or buggy provider could pin memory.

Suggested fix

In classify():

self.buffer += chunk
if len(self.buffer) > self.MAX_BUFFER_SIZE:
    overflow = self.buffer[:self.MAX_BUFFER_SIZE]
    self.buffer = self.buffer[self.MAX_BUFFER_SIZE:]
    logger.warning("ContentClassifier buffer exceeded MAX_BUFFER_SIZE; truncating")
    # flush overflow as "others"

Either enforce the cap or delete the constant and the "DoS protection" claim in the docstring.

Category: B (error handling / resource exhaustion). Severity: Medium.

Start in backend/utils/content_classifier_utils.py, especially ContentClassifier.classify() and the MAX_BUFFER_SIZE declaration. Trace how self.buffer and _process_non_tag_content handle long unrecognized input, then enforce the documented cap or remove the unused constant and DoS-protection claim. Done means the buffer cannot grow beyond the intended limit and the chosen behavior is covered by the existing classifier flow.