fix(learn): withhold label-tracking signals from the v1 classifier; leak audit

Maintainer decision "Neutralise both for v1".

1. Phantom candidates. A pair side's candidates are now only the functions that
   side declares (plane/harvest.py `declares`, the excerpt builder's
   function_bounds); a labelled function absent from a side is listed under
   `absent_on_side` and never asked, and a side declaring none of its verify
   family's functions is no verify unit. The recorded harvest units held 58
   phantom verify candidates (53 fixed, 5 vulnerable sides). build_features.py
   skips label rows whose side does not declare the function and drops their
   earlier rows (MemoryStore.drop_rows); rebuilt the pair records from the
   recorded signals, no model call: 126 label rows absent_on_side (117 fixed,
   9 vulnerable) -- access_control 13, config_secrets 4, deserialization 20,
   injection 28, memory 4, output_encoding 16, path 15, untrusted_destination
   26 -- so 252 feature rows dropped (1712 -> 1460); no other row changed.

2. Engine withheld. Input profiles in learn/schema.py (`v1` withholds the
   engine instrument, `full` withholds nothing); ProgramSpec.inputs defaults to
   v1 and compile.toml records it. Neither the engine's state nor its values
   reach the candidate, examples, demonstrations or the instruction proposer's
   error summary, and the Gower retrieval distance leaves the engine block out.
   Stored records are unchanged. design.md Risks records the decision and that
   the engine returns only through a pre-registered A/B on assumed-benign and
   benign-control negatives.

3. `ousast learn audit-leaks` (learn/leaks.py): per feature and per instrument
   state, separation = (pairs the signal alone puts on the right side - on the
   wrong side) / pairs, over all pairs and per family; flagged at >= 0.20 over
   >= 10 pairs. Counts under
   benchmarks/measurements/2026-10-01-decision-engine-leak-audit/ (before and
   after the rebuild). After: facts.function_lines (0.80 untrusted_destination,
   0.75 config_secrets, 0.26 injection; higher on fixed), verify.turns_mean
   0.67 config_secrets, eng.* 0.35-0.40 output_encoding, verify.flag_a/votes/
   site_offset and agree.final 0.20-0.30 on 10-12 pair families. Before the
   rebuild the phantom fixed sides drove state.source/state.roles/ms.flagged
   to 1.0 and state.engine to 0.95 on deserialization.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01EmM1j46qgKMh3zqD1QMwEo

src/openultrasast/learn/cli.py

@@ -53,8 +53,19 @@ def add_commands(learn_sub: Any) -> None:
         if name == "curve":
             sub.add_argument("--subset", type=int, default=200)
 
+    leaks = learn_sub.add_parser("audit-leaks", help="how well each signal alone separates the sides of a pair (no model)")
+    leaks.add_argument("--memory", help="the store (default OUSAST_MEMORY, then results_root()/memory)")
+    leaks.add_argument("--units", type=Path, help="the harvest units.json (unit -> case and side) for feature rows without a side")
+    leaks.add_argument("--profile", choices=("static", "plane"), default="plane")
+    leaks.add_argument("--kind", default="features", help="the row kind to audit (features, or example)")
+    leaks.add_argument("--threshold", type=float, help="flag a separation at least this (default 0.20)")
+    leaks.add_argument("--min-pairs", type=int, help="over at least this many pairs (default 10)")
+    leaks.add_argument("--out", type=Path, help="write the counts-only report here")
+
 
 def run(args: argparse.Namespace) -> int:
+    if args.learn_command == "audit-leaks":
+        return _audit_leaks(args)
     if args.learn_command == "memory" and args.memory_command == "build":
         return _memory_build(args)
     if args.learn_command == "memory" and args.memory_command == "embed":
@@ -225,4 +236,31 @@ def _memory_build(args: argparse.Namespace) -> int:
     return 0
 
 
+def _audit_leaks(args: argparse.Namespace) -> int:
+    """Exit 2 when the store holds no pair to audit (an unread store is not a clean one)."""
+    from ..plane.memory import open_store
+    from .leaks import MIN_PAIRS, THRESHOLD, audit, pair_rows, sides_from_units
+
+    store = open_store(args.memory)
+    rows = [r.row for r in store.rows(kind=args.kind)]
+    units = json.loads(args.units.read_text(encoding="utf-8")) if args.units is not None else {}
+    pairs, skipped = pair_rows(rows, sides_from_units(units), args.profile)
+    if not pairs:
+        print(f"learn audit-leaks: {len(rows)} {args.kind} rows in {store.describe()}, no pair to audit {skipped}", file=sys.stderr)
+        return 2
+    threshold = THRESHOLD if args.threshold is None else args.threshold
+    report = audit(pairs, args.profile, threshold=threshold, min_pairs=MIN_PAIRS if args.min_pairs is None else args.min_pairs)
+    report = {"rows_read": len(rows), "kind": args.kind, "skipped": skipped, **report}
+    text = json.dumps(report, indent=1, sort_keys=True) + "\n"
+    if args.out is not None:
+        args.out.parent.mkdir(parents=True, exist_ok=True)
+        args.out.write_text(text, encoding="utf-8")
+    print(json.dumps({"rows_read": len(rows), "pairs": report["pairs"], "flagged_signals": report["flagged_signals"]}, sort_keys=True))
+    for cell in report["flagged"]:
+        print(
+            f"  {cell['separation']:.3f}  {cell['signal']}  [{cell['scope']}]  {cell['correct']}-{cell['wrong']} of {cell['pairs']} pairs"
+        )
+    return 0
+
+
 __all__ = ["add_commands", "run"]