src/verifiable_rl_coder/rewards/runtime.py

"""Secondary reward — faster code scores higher, bounded `[0, 1]`.

Scales the sandbox's measured `runtime_ms` against a budget:

    R = max(0, (budget_ms - runtime_ms) / budget_ms)

`runtime_ms = 0`  → 1.0 (ideal)
`runtime_ms = budget_ms` → 0.0
`runtime_ms > budget_ms`   → 0.0 (clamped)

Program failures (timeout, OOM, runner error) always return 0.0 — a program
that didn't complete has no meaningful runtime to reward.

Like `lint_reward`, this is a TIEBREAKER. Weight it small in the composite.
"""

from __future__ import annotations

from ..sandbox.runner import RunResult

DEFAULT_BUDGET_MS: int = 5000


def runtime_reward(result: RunResult, *, budget_ms: int = DEFAULT_BUDGET_MS) -> float:
    """Return reward in `[0, 1]` — faster runtime → higher score."""
    if result.timed_out or result.oom or result.error is not None:
        return 0.0
    if budget_ms <= 0:
        return 0.0
    return max(0.0, (budget_ms - result.runtime_ms) / budget_ms)