feat(correlation): establish entity-resolution baseline and acceptance dataset

## Parent

#5981

## What to build

Create the evidence package that determines whether exact entity clustering fixes real user-visible correlation gaps without damaging existing cards.

## Acceptance criteria

- [ ] Capture at least 500 representative signal labels across at least three days, with a reproducible capture method and documented provenance/licensing posture.
- [ ] Report current cluster membership, dropped signals, key distribution, and p50/p95 runtime.
- [ ] Hand-label a random 100 residual drops into no-entity, missing-alias, fuzzy-variant, and unregistered-entity classes, then record the Phase 1b gate result.
- [ ] Build a held-out benchmark of at least 100 multi-entity labels with a labeled primary entity, including explicit WTI and Brent cases.
- [ ] Pre-register the affected user surface, 28-day volume, primary outcome metric, baseline, minimum detectable lift, measurement window, and continue-or-disable rule.

## Blocked by

None — can start immediately.

Start by reading parent issue #5981 and the correlation work it describes; no source files or tests are named here. Define the reproducible capture and labeling process, then produce the requested baseline, held-out benchmark, and pre-registered measurement plan so every acceptance criterion is recorded.

Written by the indexing model from the issue text.