feat(admin): prune_dataset() scoring module — rfc-corpus-growth-controls (task 1/9)

CHANGELOG.md

@@ -62,6 +62,7 @@ version numbers (so `0.1.0b1` is the first beta of the `0.1.0` line).
 
 ### Added
 
+- feat(admin): new `prune_dataset()` module + scoring rubric (`corpus_forge/admin/prune.py`) — first step of `rfc-corpus-growth-controls`. CLI verb + GrowthConfig block land in follow-up RFC tasks. Postgres / SQLite dispatch goes through a small `_is_postgres_like` capability probe (`_paramstyle == "pyformat"` first, class-name `"postgres"` substring as fallback) so we don't lean on a single brittle name check; SQLite branch chunks the IN-list at `_SQLITE_BATCH_SIZE = 500` ids. `PruneReport.duplicate_density_available` exposes whether the MinHash quality signal ran (promoted off the head candidate's `sub_scores` so every element of `selected` is now shape-uniform). Named-but-unknown datasets raise `ValueError` before any candidate walk — critical safety guard under `apply=True` so a typo'd name can never delete from the wrong scope. 22 unit tests in `tests/unit/test_prune_scorer.py` (up from 17 in the initial round) lock the rubric, the dispatch heuristics, both delete paths, and the unknown-dataset refusal.
 - `tests/unit/test_cli_human_friendly.py` — first two tests against
   the human-friendly CLI testable properties: (1) doctor's
   `_check_config_present` pins the `corpus-forge setup` recovery

corpus_forge/admin/__init__.py

@@ -24,4 +24,10 @@
 
 from __future__ import annotations
 
-__all__: list[str] = []
+from corpus_forge.admin.prune import PruneCandidate, PruneReport, prune_dataset
+
+__all__: list[str] = [
+    "PruneCandidate",
+    "PruneReport",
+    "prune_dataset",
+]