feat(research): exclusivity enrichment harness (selection-learning Commit 2, blind to output)

Build-only to the enrichment LOCK (2026-06-24). New module
selection_learning_enrich.py with its own CLI (--enrich / --enrich-preflight) and
+18 tests; no code added to byte-capped selection_learning.py.

- E1 causal leg-completeness / exclusivity feature (pivot-structural, k*=3),
  computed via a second per-endpoint truncated detect (build_exclusivity), same
  pattern as build_retro_features. Distinct from cleanliness by construction.
- E2 nested baseline = the CURRENT Stage-2 model (live k<=3 set); enriched = +exclusivity.
- E3/E4 reuse build_candidates / fit_logreg / AP / decision_point_bootstrap verbatim;
  4h k=3 primary, 1M/1w/1d context; blind verdict enrichment_helps /
  no_enrichment_signal / enriched_worse_check. Distinctness corr + exclusivity
  distribution reported (not gates).

No run executed here; no enriched AP/lift computed. Run needs frozen-data parity
(--enrich-preflight first; do NOT data.fetch --refresh). Gates green (ruff/format,
601 pytest, cov 74%).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: JohnCCarter