Accurate forecasting of yield is critical for commercial strawberry production, where daily harvest estimates directly influence labour allocation, logistics planning, and market supply commitments. Nevertheless, farm-scale forecasting operates under significant constraints, including limited multi-year history, seasonal discontinuities, seasonal zero-yield boundaries, and heterogeneous operational data sources. While most existing forecasting studies are evaluated on extensive datasets, this work specifically investigates yield prediction performance under limited-data conditions. We employ a leakage-aware, biologically informed framework for daily strawberry yield forecasting using six consecutive production years (2020–2025) from a commercial UK farm. The modelling pipeline integrates rolling environmental summaries and multi-horizon yield lag features while preserving seasonal boundaries and enforcing strict temporal alignment. A comprehensive benchmarking study evaluates linear models, distance-based methods, tree ensembles, gradient boosting frameworks, neural architectures, and a tabular foundation model (TabPFN) under three validation approaches: random train–test splitting, Leave-One-Out Cross-Validation, and forward-year generalisation. The results show that the evaluation design materially affects the apparent performance. Under deployment-relevant forward-year validation (train 2020–2024, test 2025), TabPFN achieves MAE of 438.53 and coefficient of determination (R2) of 0.696, outperforming classical baselines. Tree-based ensembles also exhibit strong generalisation, whereas neural sequence models trained from scratch underperform in the limited-data. The findings indicate that biologically motivated temporal features capture substantial predictive structure and that robust inductive bias is more beneficial than architectural complexity in small seasonal datasets.