Module height_predictor 

Bayesian height prediction with conformal bounds for virtualized lists. Bayesian height prediction with conformal bounds for virtualized lists.

Predicts unseen row heights to pre-allocate scroll space and avoid scroll jumps when actual heights are measured lazily.

Mathematical Model

Bayesian Online Estimation

Maintains a Normal-Normal conjugate model per item category:

Prior:     μ ~ N(μ₀, σ₀²/κ₀)
Likelihood: h_i ~ N(μ, σ²)
Posterior:  μ | data ~ N(μ_n, σ²/κ_n)

where:
  κ_n = κ₀ + n
  μ_n = (κ₀·μ₀ + n·x̄) / κ_n
  σ²  estimated via running variance (Welford's algorithm)

Conformal Prediction Bounds

Given a calibration set of (predicted, actual) residuals, the conformal interval is:

[μ_n - q_{1-α/2}, μ_n + q_{1-α/2}]

where q is the empirical quantile of |residuals|. This provides distribution-free coverage: P(h ∈ interval) ≥ 1 - α.

Failure Modes

Condition	Behavior	Rationale
No measurements	Return default height	Cold start fallback
n = 1	Wide interval (use prior σ)	Insufficient data
All same height	σ → 0, interval collapses	Homogeneous data
Actual > bound	Adjust + record violation	Expected at rate α

Structs

HeightPrediction

A prediction with conformal bounds.

HeightPredictor

Bayesian height predictor with conformal bounds.

PredictorConfig

Configuration for the height predictor.