节点 n16

Plateau at ds4–ds6. Shipping ds4-equivalent via principled auto-scaling (2×horizon/gap). Finalizing code:

ntime: per-type PCA on many types may exceed budget — limit to types with ≥20 cells in both stages, skip others. Engineer should run one small config first (2 types, 500 cells) to validate before full scan.",
  "family_id": "composition_program",
  "mechanism": "Decompose population change into (1) type-proportion shift via shrunk log-ratios and (2) within-type expression program estimated in PCA space with limma-style empirical Bayes shrinkage per principal component; synthesize by resampling real cells at new proportions and applying the shrunk displacement.",
  "vs_constant_shift": "Constant shift applies one global or per-type vector uniformly to all genes. Here: (a) composition and within-type parts are estimated and toggled independently; (b) within-type displacement is estimated in 20-dim PCA then back-projected, so only coordinated multi-gene programs survive; (c) EB shrinkage gives per-PC shrinkage proportional to estimation uncertainty (fewer cells → more shrinkage), not a single global alpha; (d) zero-pattern of each cell is preserved (displacement only at non-zero positions).",
  "mechanism_evidence": "Engineer should report: (1) number of types with non-zero shrunk composition change and their direction; (2) per-type L2 norm of the shrunk within-type displacement vs raw delta (shrinkage ratio); (3) four-group decomposition for 'composition-only', 'within-type-only', and 'both' variants; (4) de_score and de_direction raw values must move (not just mmd_u); (5) zero-fraction and per-gene variance of output vs input to confirm no densification.",
  "mechanism_off_control": "Off-control: set K_comp=1e9 (composition shrinkage → 0, proportions unchanged) AND s0^2=1e9 (EB prior → ∞, within-type displacement → 0). With both off, the program reduces to resampling stage-2 cells at their original proportions with no displacement = equivalent to copy_last of stage 2. Output md5 should match a stage-2 subsample. Run this as first sanity check.",
  "sources": []
}
```