PSTAR: Popularity Stranding and Training-Free Alignment & Repair for Tail Forgetting in User Foundation Models

Initializing item embeddings from a pretrained language model and fine-tuning them on user interactions is standard practice in large-scale recommendation systems. Following this recipe at a quick-commerce platform, we train a user foundation model on six months of interaction sequences with an unfrozen item table. While unfreezing lifts aggregate recall by **51%**, standard volume-weighted evaluations conceal a critical flaw: unfreezing silently collapses retrieval performance for long-tail items by **$13\times$**. We diagnose three core properties of this failure: **(i)** stranded tail items remain structurally intact (median cosine $0.975$ to initialization); the damage is positional rather than destructive, a failure mode we term *popularity stranding*; **(ii)** forgetting is deterministic: distinct random seeds strand the same events (Jaccard $0.997$); and **(iii)** the causally sound training fix is *outperformed* by a post-hoc repair under real-world adoption constraints. That repair is **PSTAR** (Popularity Stranding and Training-Free Alignment and Repair), a **training-free**, closed-form method that realigns stranded tail items via a local $k$-nearest-neighbor displacement field during index compilation in seconds. Extending by construction to unseen items, PSTAR restores zero-interaction tail recall to baseline levels (achieving **80%** of the performance of a reference model trained on **$10\times$** data) while preserving **98%** of head gains and returning **14%** of the catalog to the retrievable pool.