agent functions: one failure contract for the OpenEnv, NeMo Gym and Harbor legs

[nblintao]

Stacked on #2801.

Purpose

The OpenEnv, NeMo Gym and Harbor agent functions ended failed episodes three different ways that all trained as reward 0 without saying why — and OpenEnv's comment claimed returning None discards the sample, which it never did. Give the three legs one failure contract.

What

One rule, applied in place in each leg (no shared module yet — this makes the copies identical so a later extraction is mechanical):

• Discard (InfraAbort) only what the policy cannot have caused: sandbox create failed past the retry budget (SandboxUnavailable), agent/env server unreachable (ServerUnreachable), a tbench2 server without the canonical scoring contract (NonCanonicalVerifier).
• Score 0 with a named exit_status for everything else, wall-clock timeouts included (TimeLimitExceeded) — a discarded timeout is an exit the policy can learn to take.
• One exit_status vocabulary across the legs (Harbor's set plus the three discard causes), documented next to the who-can-cause-what table.

Behaviour change

• New discards: an unreachable agent/env server after the policy started, and OpenEnv episodes scored by a non-canonical server — both previously trained as false-negative 0s.
• Samples that were already reward 0 stay reward 0, now with an exit_status the dashboards can split on.
• OpenEnv's exit_status strings change (timeout → TimeLimitExceeded, completed → Submitted).