feat: [no-ticket] harbor - transfer targets (per-target executor-model override at finalize)

vero/src/vero/evaluation/engine.py

@@ -193,22 +193,39 @@ async def evaluate_admin(
         split: str,
         commit: str,
         sample_ids: list[int] | None = None,
+        model: str | None = None,
     ) -> Experiment:
         """Admin/verifier evaluation: explicit ``task``, no budget, no allowlist.
 
         Unlike :meth:`evaluate` (which is bound to ``default_task`` and metered),
         this scores an arbitrary ``(task, dataset_id, split)`` — including held-out
         tasks/splits the agent never had access to. Used by the verifier to score
         the selected commit on its configured targets.
+
+        ``model`` overrides the executor model for this one eval (rides
+        ``task_params`` so the eval strategy can honor it; the Mode-B
+        HarborRunner does, the Mode-A vero-task path ignores it). Used for
+        transfer targets: scoring the champion under a model it was NOT
+        optimized on.
         """
+        params = self.run_constraints
+        if model is not None:
+            params = params.model_copy(
+                update={
+                    "task_params": {
+                        **(params.task_params or {}),
+                        "harbor_model_override": model,
+                    }
+                }
+            )
         return await self.evaluator.evaluate(
             commit=commit,
             dataset_id=dataset_id,
             split=split,
             task=task,
             sample_ids=sample_ids,
             db=self.db,
-            evaluation_parameters=self.run_constraints,
+            evaluation_parameters=params,
         )
 
     def status(self) -> dict[tuple[str, str], SplitBudget]:

vero/src/vero/harbor/build/compiler.py

@@ -236,6 +236,7 @@ def _serve_config(
                 "split": t.split,
                 "reward_key": t.reward_key,
                 "sample_ids": t.sample_ids,
+                "model": t.model,
             }
             for t in config.targets
         ],

vero/src/vero/harbor/build/config.py

@@ -36,6 +36,12 @@ class TargetSpec(BaseModel):
     split: str
     reward_key: str = "reward"
     sample_ids: list[int] | None = None
+    # Executor-model override for this target (transfer probe; Mode B only):
+    # score the selected commit AND the baseline under a model it was not
+    # optimized on, so model-specific couplings (measured live: hardcoded
+    # temperature=0 crashing 72/72 on an executor that rejects it) surface at
+    # finalize instead of one substrate away.
+    model: str | None = None
 
 
 class _BuildConfigBase(BaseModel):

vero/src/vero/harbor/config.py

@@ -40,6 +40,32 @@ class HarborConfig:
     # without running anything). None keeps the current behavior: the
     # candidate env supplies harbor, and is trusted to.
     harbor_requirement: str | None = None
+    # Bounded within-eval retry for infra-destroyed samples. A sample whose
+    # EVERY attempt died of a transient infrastructure cause (connection,
+    # timeout, rate limit, 5xx) was never measured at all: re-run it after a
+    # backoff instead of booking the outage as a permanent error. Measured
+    # live: a 65-second host DNS blip killed 44 of 72 attempts of one eval
+    # with ConnectionError, and nothing in the record distinguished the blip
+    # from a bad candidate.
+    #
+    # OFF BY DEFAULT, and it must stay off when the candidate is an
+    # adversarial optimizer. The qualifying predicate is built from exception
+    # types raised inside candidate code, and agents are stochastic: a
+    # candidate that raises an allowlisted exception whenever an attempt is
+    # going badly loses nothing on partially-good samples (its fakes
+    # zero-fill like honest failures) but converts every all-bad sample from
+    # a booked 0.0 into a fresh re-roll. That is one-sided selection over
+    # attempt sets, exactly what the zero-fill invariant exists to prevent.
+    # Enable only for trusted-candidate evaluations (frozen agents,
+    # operator-run matrices), where re-measuring an outage is pure signal
+    # recovery. Candidate crashes and exhausted key budgets never retry
+    # regardless (a crash is a result; a spent key cannot recover by
+    # waiting), and recovered samples carry an ``infra_retry`` audit marker.
+    infra_retry_rounds: int = 0
+    # Backoff before retry round N is N times this many seconds. Transient
+    # infra needs time, not immediacy: instant retries burned 6 of 8 run
+    # attempts inside one live DNS blip.
+    infra_retry_delay_s: float = 30.0
     extra_args: list[str] = field(default_factory=list)  # passthrough harbor run flags
 
     def __post_init__(self) -> None:
@@ -50,6 +76,18 @@ def __post_init__(self) -> None:
                 f"aggregate_attempts must be 'best' or 'mean', got "
                 f"{self.aggregate_attempts!r}"
             )
+        if self.infra_retry_rounds < 0:
+            raise ValueError(
+                f"infra_retry_rounds must be >= 0, got {self.infra_retry_rounds}"
+            )
+        # A zero (or negative) delay silently nullifies the backoff, and an
+        # instant retry re-enters the same outage: 6 of 8 run attempts once
+        # burned inside a single live DNS blip for exactly this reason.
+        if self.infra_retry_rounds > 0 and self.infra_retry_delay_s <= 0:
+            raise ValueError(
+                f"infra_retry_delay_s must be > 0 when infra retries are "
+                f"enabled, got {self.infra_retry_delay_s}"
+            )
 
     @property
     def is_registry(self) -> bool: