LLM Failure-Mode Taxonomy

**Role:** AI safety researcher building the team's shared vocabulary for LLM bugs.

**Context:** Team is shipping LLM features fast and producing inconsistent bug reports. Engineering and QA don't have a shared vocabulary for "what went wrong."

**Task:** Build the taxonomy:
1. **Hallucination** types: factual confabulation, citation invention, format invention.
2. **Refusal** types: over-refusal, under-refusal, miscalibrated refusal.
3. **Drift** types: persona drift, format drift, scope drift.
4. **Reasoning** failures: shallow CoT, math errors, contradiction tolerance.
5. **Tool-use** failures: wrong tool, wrong args, ignored output.
6. **Format** failures: invalid JSON, broken markdown, encoding mismatch.
7. **Latency / cost** failures: token waste, slow tool calls, over-reasoning.
8. **Safety** failures: PII leakage, jailbreak success, copyright leak.

For each: definition, example, observable signal in logs, who's responsible for fixing.

**Constraints:**
- Every category has a CONCRETE EXAMPLE from real production.
- Each failure has a single "owner" team.
- Avoid academic terms when ops terms exist.

**Output format:** Taxonomy doc + bug-template (Jira / Linear / GitHub) using these labels.