fix(evals): derive judge model levels from the registry (claude-opus-5 was silently demoted to medium)

[elhoim]

The three model-based graders each hardcoded a judge-model → inference-level map. The lineup moved and the maps didn't, so both current top-tier IDs fell through to the ?? 'medium' default:

judge_model	resolved to	should be
claude-opus-5	medium	high
claude-sonnet-5	medium	medium (right only by accident)

The opus case is the damaging one: an eval configured with the strongest judge silently ran on Sonnet-tier inference, and nothing surfaced the demotion — the grader still returned scores, just from a weaker model than asked for.

The fix

Derive the current lineup from EFFORT_MODEL/CURRENT in models.ts rather than restating it. That file already describes itself as the single edit point on a model release, and its header cites exactly this failure (ContextAudit checking for claude-opus-4-7 long after claude-opus-4-8 shipped) as the reason it exists.

const levelMap: Record<string, InferenceLevel> = {
  ...legacyLevels,
  ...Object.fromEntries(
    Object.entries(EFFORT_MODEL).map(([level, tier]) => [CURRENT[tier], level as InferenceLevel]),
  ),
};

Superseded IDs (claude-opus-4-8, claude-opus-4-6, claude-sonnet-4-6, and the two dated 2025 IDs) stay in a small legacyLevels table so existing eval configs keep resolving. claude-fable-5 and claude-haiku-4-5-20251001 came out of the hardcoded list because they're now produced by the derivation with identical values.

A future model bump flows through automatically instead of stranding a judge here.

Verified

Resolution asserted against the shipped expression:

• claude-opus-5 → high, claude-sonnet-5 → medium, claude-fable-5 → max, claude-haiku-4-5-20251001 → low
• all five legacy IDs unchanged
• unknown model and empty string still fall back to medium

All three graders transpile clean.

Companion to #1656, which bumped CURRENT.opus to claude-opus-5 — that bump is what made the stale maps observable.