# Sprint 118 Plan: Distributed Failure Capture and Evidence Normalization

## Context

Sprint 117 connected iteration execution to HiveMind job dispatch. The next gap is
consistent failure evidence capture from distributed jobs so downstream debugging
is deterministic instead of ad-hoc.

---

## Goals

1. Normalize distributed job outputs into deterministic failure evidence
2. Preserve replay-ready execution context with stable schemas
3. Add MCP tools to capture and query distributed failure evidence
4. Keep evidence machine-readable and deterministic across reruns

---

## Steps

### Step 1419: `DistributedFailureEvidence` schema + normalizer (12 tests)
### Step 1420: Job-output canonicalization and truncation utilities (10 tests)
### Step 1421: Distributed repro command resolver model (10 tests)
### Step 1422: Evidence bundle store + deterministic ordering (10 tests)
### Step 1423: `whetstone_capture_distributed_failure` MCP tool (8 tests)
### Step 1424: `whetstone_list_distributed_failures` MCP tool (8 tests)
### Step 1425: `whetstone_get_distributed_failure_bundle` MCP tool (8 tests)
### Step 1426: Distributed evidence quality scorer model (8 tests)
### Step 1427: Distributed evidence packet export model (8 tests)
### Step 1428: Sprint 118 integration summary + regression (8 tests)

---

## Architecture Gate

- Failure evidence packets must be deterministic for identical inputs
- Bundle ordering must be stable and replayable
- Tool outputs must avoid free-text-only ambiguity
- Max 600 lines per header