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/compile.py

@@ -52,7 +52,7 @@ from .program import (
     ProgramSpec,
     label_word,
 )
-from .schema import SCHEMA_VERSION, feature_set_digest
+from .schema import BY_NAME, DEFAULT_INPUTS, SCHEMA_VERSION, feature_set_digest, withheld
 
 DEFAULT_SPEC = Path(__file__).resolve().parent / "compile.toml"
 KIND = "llm-program"
@@ -152,8 +152,13 @@ class Bootstrap:
         self.rejected[why] = self.rejected.get(why, 0) + 1
 
 
-def _signal_summary(example: Example) -> dict[str, Any]:
-    ran = {k: v for k, v in example.x.items() if v not in (None, 0, False) and not k.startswith("lang.")}
+def _signal_summary(example: Example, inputs: str = DEFAULT_INPUTS) -> dict[str, Any]:
+    """The non-zero signals of a mistake for the instruction proposer, without the instruments ``inputs`` withholds."""
+    hidden = withheld(inputs)
+    ran = {
+        k: v for k, v in example.x.items()
+        if v not in (None, 0, False) and not k.startswith("lang.") and not (k in BY_NAME and BY_NAME[k].instrument in hidden)
+    }  # fmt: skip
     return dict(sorted(ran.items())[:12])
 
 
@@ -179,7 +184,7 @@ def bootstrap(
         right = decision.verdict == label_word(example.label)
         if not right:
             out.errors.append({"family": example.family, "label": label_word(example.label), "verdict": decision.verdict,
-                               "confidence": answer.confidence, "signals": _signal_summary(example)})  # fmt: skip
+                               "confidence": answer.confidence, "signals": _signal_summary(example, program.spec.inputs)})  # fmt: skip
             out.reject("wrong or unsure")
             continue
         if answer.parse_failed or answer.confidence < min_confidence:
@@ -370,6 +375,7 @@ def compile_program(
     base = ProgramSpec(
         profile=profile, model=str(classify.get("model", "deepseek-flash")), k=int(classify.get("k", 1)), k_ret=int(retrieval.get("k", 6)),
         lam=float(retrieval.get("lam", 0.5)), n1=int(retrieval.get("n1", 40)), priors=str(retrieval.get("priors", "off")),
+        inputs=str(classify.get("inputs", DEFAULT_INPUTS)),
         max_output_tokens=int(classify.get("max_output_tokens", 300)),
     )  # fmt: skip
     from .evaluate import candidate_from_example
@@ -434,6 +440,8 @@ def compile_program(
             "temperature": [0.0, 0.7],
             "confidence": "vote_mean",
             "max_output_tokens": base.max_output_tokens,
+            "inputs": base.inputs,
+            "withheld": sorted(withheld(base.inputs)),
         },
         "calibration": None,
         "operating_points": None,
@@ -491,6 +499,7 @@ def spec_of(artifact: Mapping[str, Any]) -> ProgramSpec:
         lam=float(retrieval["lam"]),
         n1=int(retrieval["n1"]),
         priors=str(retrieval.get("priors", "off")),
+        inputs=str(classify.get("inputs", DEFAULT_INPUTS)),
         max_output_tokens=int(classify.get("max_output_tokens", 300)),
     )