files per category: {'advanced': 34, 'deployment': 9, 'how-to': 12, 'reference': 24, 'tutorial': 51}
train files: 96, test files: 34
train snippets: 1550, test snippets: 637
train snippets per category: {'advanced': 492, 'deployment': 96, 'how-to': 114, 'reference': 54, 'tutorial': 794}
test snippets per category: {'advanced': 109, 'deployment': 177, 'how-to': 28, 'reference': 6, 'tutorial': 317}

accuracy=0.600 macro_f1=0.427  (baseline if always predicting majority class 'tutorial': 0.498)

confusion matrix (rows=true, cols=predicted), classes: [np.str_('advanced'), np.str_('deployment'), np.str_('how-to'), np.str_('reference'), np.str_('tutorial')]
         advanc  deploy  how-to  refere  tutori
 advance     40       1       2       1      65
 deploym     32      86       0       1      58
  how-to     12       0       2       1      13
 referen      2       0       0       3       1
 tutoria     50       4       7       5     251

calibration (predicted max-probability bucket vs observed accuracy):
  confidence [0.2,0.4): n=  57 mean_confidence=0.363 observed_accuracy=0.368
  confidence [0.4,0.6): n= 249 mean_confidence=0.505 observed_accuracy=0.510
  confidence [0.6,0.8): n= 214 mean_confidence=0.698 observed_accuracy=0.673
  confidence [0.8,1.0): n= 117 mean_confidence=0.883 observed_accuracy=0.769

median single-snippet latency (tfidf transform + logreg predict_proba): 0.628ms over 300 calls, min=0.575ms max=0.970ms

for reference only (not measured, arithmetic on published price): routing these 637 test snippets through Jev at its published input rate would cost about $0.0014 for ~33,595 approx tokens (word-count-based estimate, not a real tokenizer). This local classifier: $0, no network call.