ops(correlation): validate and activate exact entity clustering

## Parent

#5981

## What to build

Produce the release evidence for exact entity clustering and, only with separate operator approval, activate it while retaining an immediate rollback to legacy behavior.

## Acceptance criteria

- [ ] Compare all existing clusters and emitted cards before/after, reporting gained, lost, and moved signals separately.
- [ ] Require every existing cluster to retain at least 80% of its members and explain any card loss, threshold change, score change, or signal-type loss.
- [ ] Audit newly formed clusters with zero same-entity-but-unrelated-event false convergences.
- [ ] Reach at least 95% primary-entity accuracy on the held-out multi-entity benchmark with zero WTI/Brent errors.
- [ ] Keep p95 runtime below the correlation engine's 100 ms target and prove fail-closed rollback.
- [ ] Record separate activation approval, observe the pre-registered user-outcome window, and return to `legacy` if the continue rule fails.

## Blocked by

- #5984
- #5986

Start by reading parent issue #5981 and the blocking issues #5984 and #5986; the payload names no source files or tests. Done means producing before/after cluster and card evidence, meeting the accuracy and runtime thresholds, proving fail-closed rollback, and obtaining separate activation approval before observing the outcome window.

Written by the indexing model from the issue text.