fix(bayesian): surface skipped optimization and clamp warm start to search space

obliteratus/bayesian_optimizer.py

@@ -41,6 +41,56 @@ if TYPE_CHECKING:
 logger = logging.getLogger(__name__)
 
 
+# Kernel search space, declared once. `objective` suggests from these bounds and
+# the warm-start trial is clamped to them; keeping two copies is what let the
+# warm start drift outside the space it was seeding (see `_clamp_to_space`).
+KERNEL_SPACE: dict[str, tuple[float, float]] = {
+    "attn_max_weight": (0.5, 1.0),
+    "attn_peak_position": (0.1, 0.9),
+    "attn_min_weight": (0.0, 0.3),
+    "attn_spread": (0.1, 0.6),
+    "mlp_max_weight": (0.3, 1.0),
+    "mlp_peak_position": (0.1, 0.9),
+    "mlp_min_weight": (0.0, 0.3),
+    "mlp_spread": (0.1, 0.6),
+}
+
+
+def _skip(pipeline: AbliterationPipeline, reason: str, message: str) -> None:
+    """Record and surface a no-op so it cannot be mistaken for a completed run.
+
+    Reported three ways because each reaches a different audience: the stdlib
+    logger for anyone capturing logs, the pipeline log for the console panel the
+    operator is watching, and an attribute the caller persists into checkpoint
+    metadata so the provenance of an already-written model still shows it.
+    """
+    logger.warning(message)
+    log = getattr(pipeline, "log", None)
+    if callable(log):
+        log(f"WARNING: {message}")
+    try:
+        pipeline._bayesian_skipped = reason
+    except AttributeError:  # pragma: no cover - pipeline stubs may use __slots__
+        pass
+
+
+def _clamp_to_space(params: dict[str, float]) -> dict[str, float]:
+    """Clamp warm-start values into the declared search space.
+
+    Optuna rejects an enqueued value outside a parameter's distribution with a
+    warning and does not use it, so an out-of-range warm start silently costs a
+    trial and the informed starting point it was meant to provide. This is
+    reachable in normal use: `attn_peak_position` is seeded from
+    `peak_layer / (n_layers - 1)`, which for an early peak layer falls below the
+    0.1 floor — a 36-layer model peaking at layer 2 seeds 0.057.
+    """
+    out = dict(params)
+    for name, (lo, hi) in KERNEL_SPACE.items():
+        if name in out:
+            out[name] = min(max(float(out[name]), lo), hi)
+    return out
+
+
 def _measure_refusal_rate(
     pipeline: AbliterationPipeline,
     n_prompts: int = 10,
@@ -265,13 +315,31 @@ def run_bayesian_optimization(
         import optuna
         from optuna.samplers import TPESampler
     except ImportError:
-        logger.warning(
-            "Optuna not installed — skipping Bayesian optimization. "
-            "Install with: pip install optuna"
+        # Optuna is not declared in pyproject.toml, requirements.txt or uv.lock,
+        # so a documented `uv sync --locked --extra dev` install does not have it
+        # and `--method optimized` reaches this branch. `logger.warning` alone is
+        # not enough to make that visible: the `obliterate` command never calls
+        # logging.basicConfig, so this surfaces only through logging's
+        # last-resort stderr handler, while the run panel the user actually
+        # reads is fed by `pipeline.log`. The run then completes, reports
+        # success, and writes metadata recording the method that asked for an
+        # optimization which never ran.
+        _skip(
+            pipeline,
+            "optuna-not-installed",
+            "Optuna not installed — Bayesian optimization SKIPPED. Layer "
+            "weights fall back to method defaults; results are NOT optimized. "
+            "Install with: pip install optuna",
         )
         return {}
 
     if not pipeline.handle or not pipeline._strong_layers:
+        _skip(
+            pipeline,
+            "no-strong-layers",
+            "Bayesian optimization SKIPPED: no model handle or no selected "
+            "layers to optimize over.",
+        )
         return {}
 
     model = pipeline.handle.model
@@ -401,16 +469,16 @@ def run_bayesian_optimization(
         _restore_all()
 
         # Attention kernel: 4 params
-        attn_max = trial.suggest_float("attn_max_weight", 0.5, 1.0)
-        attn_peak = trial.suggest_float("attn_peak_position", 0.1, 0.9)
-        attn_min = trial.suggest_float("attn_min_weight", 0.0, 0.3)
-        attn_spread = trial.suggest_float("attn_spread", 0.1, 0.6)
+        attn_max = trial.suggest_float("attn_max_weight", *KERNEL_SPACE["attn_max_weight"])
+        attn_peak = trial.suggest_float("attn_peak_position", *KERNEL_SPACE["attn_peak_position"])
+        attn_min = trial.suggest_float("attn_min_weight", *KERNEL_SPACE["attn_min_weight"])
+        attn_spread = trial.suggest_float("attn_spread", *KERNEL_SPACE["attn_spread"])
 
         # MLP kernel: 4 params (separate — can peak at a different layer)
-        mlp_max = trial.suggest_float("mlp_max_weight", 0.3, 1.0)
-        mlp_peak = trial.suggest_float("mlp_peak_position", 0.1, 0.9)
-        mlp_min = trial.suggest_float("mlp_min_weight", 0.0, 0.3)
-        mlp_spread = trial.suggest_float("mlp_spread", 0.1, 0.6)
+        mlp_max = trial.suggest_float("mlp_max_weight", *KERNEL_SPACE["mlp_max_weight"])
+        mlp_peak = trial.suggest_float("mlp_peak_position", *KERNEL_SPACE["mlp_peak_position"])
+        mlp_min = trial.suggest_float("mlp_min_weight", *KERNEL_SPACE["mlp_min_weight"])
+        mlp_spread = trial.suggest_float("mlp_spread", *KERNEL_SPACE["mlp_spread"])
 
         # Float direction index (cross-layer interpolation, Heretic-style)
         dir_idx = trial.suggest_float("dir_idx", 0.0, max(n_layers_with_dirs - 1, 0.0))
@@ -521,7 +589,7 @@ def run_bayesian_optimization(
             "mlp_spread": 0.3,
             "dir_idx": 0.0,
         }
-    study.enqueue_trial(warm_params)
+    study.enqueue_trial(_clamp_to_space(warm_params))
 
     pipeline.log(f"Bayesian optimization: running {n_trials} trials (parametric kernel)...")
     study.optimize(objective, n_trials=n_trials, show_progress_bar=False)