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

.kiro/specs/learned-decision-engine/design.md

@@ -1163,6 +1163,15 @@ start only after the maintainer confirms their budget against the balance read j
   near-duplicate drop at cosine >= 0.98 in evaluation, and `test_boundary`. Demonstrations from the compile split
   are, by construction, from repositories never evaluated. The seed for plane Runs filters on the host, where the
   labels are; the sandbox never receives an example of its own case's group.
+- **Label-tracking signals (maintainer decision 2026-10-01: "Neutralise both for v1").** The harvest found two
+  signals that follow the pair label instead of the code: verify candidates on fixed sides named functions the side
+  no longer declares (now never produced: a side's candidates are only the functions it declares, the rest counted
+  `absent_on_side`), and the engine's "read the file, asked no question" state, which a fixed side reaches by losing
+  its sink. The v1 program withholds the engine instrument entirely -- state and values -- from prompts,
+  demonstrations, the instruction proposer and the retrieval distance (input profile `v1`, `ProgramSpec.inputs`);
+  the stored record keeps it. The engine returns only through a pre-registered A/B evaluated on assumed-benign and
+  benign-control negatives, where it cannot track a fix. `ousast learn audit-leaks` reports how well every signal
+  alone separates the sides of a pair and flags any above 0.20 (counts under `benchmarks/measurements/`).
 - **Earlier leakage paths still apply.** Hand rules and framework priors written after seeing a training repository
   (the `$wpdb` rule from PMPro, the hook source from WP Statistics, `taint.sc:911-949`): priors off by default,
   leave-one-framework-out, best-effort `taught_by = [repo]`. The mechanism exporter's train-on-test leak