fix(validator): keep agent-attributable inference declines out of no-fault

[Peyton-Spencer]

Summary

Companion to dittobench-api #120, which splits the relay's single finalize failure code apart. That code collapsed nine unrelated conditions — several entirely the harness's doing — into validator_infrastructure / model_relay_unavailable / retryable: true, the no-fault class that mints a retry grant, raises the attempt cap, and re-leases. An agent that reliably spent its own inference allowance re-leased itself forever.

One new code reaches this side, and it must not be treated as no-fault:

Code	Envelope	In _SANDBOX_INFRASTRUCTURE_CODES?	Maps to
inference_allowance_exhausted	sandbox_failure / retryable: false	No — deliberately	scoring_error

It means the harness spent the request-count or token allowance its own ticket granted, or sent one request too large to reserve (platform decline codes 4102 / 4104 / 4109). The lease was alive and the platform healthy throughout — this is the agent's.

Why there is no second code

An earlier revision of this PR added inference_lane_saturated to the set, for the other half of #120: the scorer's call spending its entire bounded backpressure budget on a platform lane that stayed full. That half is still classified as no-fault — a miner can neither provision the lane nor observe its contention, and charging it to the agent would be the mirror image of the misclassification #120 fixes — but it must not be a new code, and so it needs nothing here.

This module returns None for any code outside the set, and None becomes fail_job("scoring_error"). A brand-new no-fault code therefore charges the miner on every validator that predates the release that adds it, and the fleet lags scorer releases by design — it is on v0.36.0 / v0.34.1 today. That is the same bug #120 exists to fix, arriving through the rollout instead of through the classifier.

So #120 keeps model_relay_unavailable (in every deployed validator's allowlist since the code existed) and carries the cause in the failure's diagnostics, which this classifier never consults and therefore cannot change the class anywhere.

The asymmetry is worth naming, because it decides what is safe to ship in each direction:

• New agent codes are skew-safe. An old validator sees sandbox_failure / retryable: false and its terminal default is already the intended outcome.
• New no-fault codes are not. They need the validator side shipped and the fleet rolled first.

It is recorded in a comment next to the set, since that set is exactly where a future edit would be tempted to "finish the job".

Consequence: #288 and #120 have no ordering hazard in either direction. Neither repo depends on the other having shipped.

Why the agent code must stay out

Every code in _SANDBOX_INFRASTRUCTURE_CODES is no-fault. Adding inference_allowance_exhausted would mean an agent that reliably exhausts its own allowance re-leases itself without bound — the mnemox loop, arriving through the inference lane instead of the sandbox.

Excluded three independent ways:

1. It is not in the set.
2. The scorer sends kind: sandbox_failure, and _sandbox_infrastructure_failure_code requires validator_infrastructure before it looks at the code at all.
3. It sends retryable: false, which that function also requires to be True.

A module-level assert fails at import if the two sets ever intersect.

Tests

make lint typecheck test green — ruff clean, mypy clean on 112 source files, full suite passing.

• test_agent_attributable_inference_failures_stay_the_agents — pins the agent code to scoring_error through the poll path under both the envelope the scorer sends and the no-fault envelope it must never be smuggled under, so the guarantee holds even if a future scorer sends the wrong one.
• test_agent_inference_codes_are_never_no_fault — the set property, stated directly. A future edit adding the agent code to the infrastructure set (the natural mistake, since the neighbouring inference_lane_saturated is in it) fails here rather than in production as an unbounded re-lease loop. Also asserts the complement: if someone later moves inference_lane_saturated out, that is a decision about who pays for a saturated rail and they have to edit this line to make it.
• The same test also asserts the complement — inference_lane_saturated is not in the set and model_relay_unavailable is — so the rollout reasoning above is pinned, not just commented.
• test_unknown_sandbox_failure_code_is_not_retryable already covers unknown codes and is unchanged.

_SANDBOX_INFRASTRUCTURE_CODES collision — how it was handled

#286 (peyton/screened-image-acquisition-not-agent-fault, the companion to dittobench-api #118) edits the same set to add screened_image_unavailable. This PR is stacked on #286's branch — base is peyton/screened-image-acquisition-not-agent-fault, not main — so the two additions compose instead of conflicting, and this diff shows only its own change.

#286 was left completely untouched: its base is still main, it was not adopted into a GitHub Stack object, and no git operation was run against its branch. That branch is currently checked out in another active worktree, and stack operations against a branch held by a live worktree are exactly how restacks break.

Merge order: #286 first, then this. If #286 is abandoned, this needs a rebase onto main; the two hunks are adjacent but not overlapping, so it is a clean re-apply. There is no cross-repo ordering constraint against #120 (see above).

Also in flight and deliberately untouched: #287 (fix/idle-slots-abandon-sweep), the heartbeat v17 lease work, validator_lease_audit, the fleet-health dashboard, and the scores read path.

🤖 Generated with Claude Code

[Peyton-Spencer]

Companion: ditto-assistant/dittobench-api#120, which emits both new codes. Merge order across repos: ditto-subnet #286 → ditto-subnet #288 → (with) dittobench-api #120. This PR must not land materially after #120 in the fleet: #120 emits inference_lane_saturated, and a validator that does not recognise it would treat a saturated platform lane as terminal rather than no-fault.

[Peyton-Spencer]

Correction to the ordering note above. The earlier comment said this must not land materially after dittobench-api#120. That constraint is gone.

It existed because saturation shipped a new no-fault code that old validators would have turned into scoring_error. #120 now keeps model_relay_unavailable for saturation and puts the cause in diagnostics, so no validator needs to know anything new to keep it no-fault. The inference_lane_saturated allowlist entry has been dropped from this PR.

Remaining order is intra-repo only: #286 → #288. No cross-repo constraint against #120.