Failure surfaces and annotation-based analysis

Lesson 1.3 builds on two important foundations. In 1.1, non-determinism was quantified: the gap between what the system CAN do and what it DOES do consistently. In 1.2, the evaluation surface was mapped, covering every pipeline stage where failures could happen. But there is a gap between those two. Predictions were made about where failures COULD happen. The failures that ARE happening have not yet been systematically found. That is a significant difference. This lesson closes that gap. Raw traces (recorded end-to-end logs of user queries flowing through the AI Data Analyst v0) are examined to discover failure categories from scratch. No predefined checklist. No assumptions about what will be found. Just traces, observations, and patterns. By the end, the result is a failure taxonomy: a categorized inventory of every way the system is actually breaking, with a triage label that indicates what to do about each one.