Timeline-Weighted Contextual Bandits for Real-Time E-Commerce Intervention: Representation, Benchmarks, and an Honest Off-Policy Study

Contextual bandits are widely used for online personalization, yet many production systems compress a live shopping session into static or weakly temporal features, discarding the order, recency, and repetition structure of user behavior. We present TVCB (Timeline-Weighted Contextual Bandit), which represents each session as a compact, exponentially decayed, versioned timeline state, maintained at two timescales, that is updated in O(1) per event and consumed by low-latency bandit policies for next-best-action selection (e.g., discount, free shipping, social proof, or deliberate non-intervention). On top of this state we study a spectrum of policies: an interpretable fixed-score TVCB, learned LinUCB and posterior-covariance Thompson-sampling policies over timeline features, a discounted variant for non-stationary rewards, and a representation corral meta-bandit that learns online whether the timeline representation or a simpler cluster representation pays in a given deployment, rather than assuming either. Because no public, propensity-logged benchmark exists for session-level intervention (as opposed to item recommendation), we also release TIB (Timeline Intervention Benchmark), a versioned, byte-reproducible suite spanning (i) a hard synthetic tier whose rewards depend on decayed event sequence patterns, (ii) a digital-twin tier emitting randomized intervention logs with a production-style schema, exact propensities, and full counterfactual outcomes, and (iii) a semi-real tier built on RetailRocket browsing timelines, with a datasheet, machine-readable metadata, a normative evaluation protocol, and a baseline leaderboard. We evaluate with direct, inverse-propensity, and doubly robust estimators under multi-seed confidence intervals, and additionally validate the pipeline on two real datasets: the Open Bandit Dataset and the Hillstrom randomized e-mail experiment, where fixed-arm sanity identities reproduce the published treatment rates and calibrate the pipeline. Our findings are deliberately two-sided: timeline-aware learned policies dominate when intervention response depends on behavioral sequence and hesitation dynamics (e.g., +3.9% direct reward and non-overlapping confidence intervals on the hard synthetic benchmark), but cluster-level and static-context baselines remain competitive, and sometimes better, when response is driven by coarse intent alone. The corral meta-policy resolves this tension empirically: it routes the vast majority of decisions to the better representation in both regimes and lands near the per-regime best expert without prior knowledge of which regime holds. We characterize these regimes with a nine-scenario stress ablation, report an in-process decision latency of p99 ≈ 0.02 ms, and discuss what evidence would be required to claim production incrementality. All results are reproducible from committed scripts and seeds. The TIB benchmark suite is archived at https://doi.org/10.5281/zenodo.20651686.