Default to Claude Opus 5.5 with refreshed benchmarks

[judagent]

Summary

• Default advisor: anthropic/claude-opus-5-5#xhigh (was claude-fable-5-1#xhigh), with one opencode retry when the Anthropic route is unavailable. The serving route is recorded as advisorProvider on metrics rows.
• Self-consultation guard generalized from hard-coded Fable to the configured advisor model (any provider route, any effort, normalized across nested/dotted IDs). New skipped_self outcome; legacy skipped_fable history parsing kept.
• Benchmarks: fresh Artificial Analysis snapshot (673 models, fetched 2026-09-23) and 69 bindings covering Opus 5.5 (direct + opencode, all efforts), GPT-6 Sol/Luna, Grok 4.7, new Fable 5.1 / GPT-6 Astra effort levels, gateway DeepSeek V4.1 Flash routes, and Muse Spark 1.2. The generator now fails on missing slugs instead of skipping them, and a seed test asserts every bundled binding resolves.
• Requester variant "default" treated as unset for benchmark matching; usage-report eligibility proxy parameterized (--advisor-model, default claude-opus-5-5).

Behavioral notes

• Fable sessions (including the plan agent on claude-fable-5-1#xhigh) will now see the advisor tool; Opus 5.5 sessions are now the blocked ones.
• Opus 5.5 has no AA Coding Index yet, so coding comparisons report missing_advisor until Artificial Analysis publishes one. Watch skipped_typesafe rates for the first few days.
• deepseek/deepseek-flash and muse-spark-1.3#high stay unmapped (no AA entry; no effort substitution per policy).

Verification

• bun test: 361 pass, 0 fail; tsc --noEmit clean; bun run build clean.
• Compiled CLI benchmarks status --model: 25/27 top-usage models matched (2 intentional exclusions above); anthropic/claude-opus-5-5#xhigh matched with evaluated effort xhigh.
• Fallback covered on both discovery namespaces (legacy catalog + V2); failed fallback reports the primary reason.

Models: anthropic/claude-opus-5-5:xhigh, meta/muse-spark-1.3:max, anthropic/claude-fable-5-1:xhigh