files: 151 total, 120 train, 31 test
pairs: 2041 train, 1187 test
token-space vocabulary: 150 words, e.g. ['fastapi', 'use', 'py', 'docs_src', 'hl', 'path', 'python', 'using', 'https', 'example']

task A zero-shot baseline (context TF-IDF used directly as query, no predictor trained):
  hit@1=0.005 hit@5=0.127 MRR=0.068  (n=1187, random-guess hit@1 would be 0.0008)

task A: embedding-space prediction (Ridge, context TF-IDF -> target TF-IDF)
  frac  n_train   hit@1   hit@5     MRR  train_s
  0.10      204   0.011   0.026   0.025     0.05
  0.25      510   0.012   0.051   0.039     0.19
  0.50     1020   0.018   0.061   0.049     0.29
  1.00     2041   0.020   0.083   0.062     0.47

task B: token-space prediction (LogisticRegression x150 words, context TF-IDF -> word presence)
  frac  n_train  micro_F1  train_s
  0.10      204     0.001     2.19
  0.25      510     0.105     0.64
  0.50     1020     0.120     0.97
  1.00     2041     0.230     1.64

task B majority-class baseline (predict train-set >50%-frequent words every time): micro_F1=0.000

sample efficiency, 100%->10% of train pairs:
  task A hit@1: 0.020 -> 0.011  (+45.8% relative change)
  task B micro_F1: 0.230 -> 0.001  (+99.4% relative change)
  task B's relative drop is 2.17x steeper than task A's

writeup ratios (full precision):
  data scale 10%->100%: 10.00x
  task A hit@5 100%/10%: 3.16x
  task A hit@1 100% vs random baseline: 24.00x
  zero-shot vs trained(100%) hit@1: trained is 4.00x zero-shot
  zero-shot vs trained(100%) hit@5: zero-shot is 1.54x trained
  zero-shot vs trained(100%) MRR: zero-shot is +10.2% relative to trained