Vehicle-Conditional Split-Conformal Calibration for Risk-Budgeted Sub-Second Proxy-Triggered Vehicle Instability Warnings from Past-Only Sensor Slices

Emergency maneuvers can drive vehicles into severe instability regimes within sub-second time scales, motivating last-moment warning interfaces with auditable false-alarm budgets. We study a proxy-triggered imminent-recognition setting: given a 0.1 s past-only slice of onboard signals, decide whether a conservative physics-defined instability proxy will trigger within the next τ=0.2 s. The contribution is, therefore, a calibrated warning for a safety-relevant surrogate event, not a claim of predicting crashes or true instability outcomes directly. Because the corpus is terminal-phase aligned, the default causal monitor (w=d=0.1 s, k=2) is warnable on only 18.3% of event runs; we, therefore, report run-level effectiveness both overall and conditional on warnability. We learn a lightweight hazard scorer and convert its scores into an operator-facing alarm rule via split-conformal calibration on held-out negative slices, exposing a slice-level false-alarm budget α with finite-sample, one-sided control of the marginal slice-level false positive rate (FPR) on exchangeable negatives. To address fleet heterogeneity, we additionally calibrate vehicle-conditioned (Mondrian) thresholds, enabling per-vehicle risk budgeting without retraining separate models. On the held-out test split at τ=0.2 s, the scorer achieves AUPRC ≈0.251 against a base rate of 0.638%, AUROC ≈0.986, and ECE ≈0.034. After calibration at α=5%, realized slice-level FPR concentrates near the prescribed budget while slice-level TPR on imminent positives remains high (≈0.982). We explicitly separate this slice-level guarantee from empirical run-level metrics such as FARrun, EWR on warnable runs, and lead time, and we report dependence and shift diagnostics to delineate where the guarantee may degrade. The reported μ-sensitivity analyses concern run-level descriptor perturbation and omission rather than validation of a within-run friction estimator with temporal lag. The result is a transparent, risk-budgeted monitoring primitive for last-moment vehicle-stability warning under clearly stated exchangeability assumptions.