[Task] question-answering EP coverage investigation

Overview

Extractive QA models (BERT, RoBERTa, DeBERTa, ELECTRA) identify answer spans from a context passage. 9 of 10 models already pass all 3 EPs. This issue targets the single remaining model with incomplete coverage.

tau/splinter-base uses the SPLINTER architecture (span-level representation) which differs structurally from standard BERT-family models — the VitisAI failure may be related to its span selection head or non-standard attention patterns.

Agent Scenarios

• See #270 for full QA agent scenarios (document Q&A, customer support, research assistant, on-device copilot)

EP Coverage Status

Reference: all other 9 QA models pass all 3 EPs.

Acceptance Criteria

• tau/splinter-base passes all 3 EPs (investigate VitisAI failure)