feat(plugins): add Claude Opus 5.5 and GPT-6 support

[SpiGAndromeda]

Summary

Add prompting and routing guidance for Claude Opus 5.5 and the GPT-6 models (Astra, Sol, Luna) to llm-author and work-orchestrator. Recalibrate work-orchestrator's per-model routing against a local replication, because five per-model GPT-5.6 observations the routing table relied on did not reproduce.

Changes

Prompting guidance (llm-author 3.13.0)

• Add an Opus 5.5 section to claude-5-guide.md: default effort medium instead of Opus 5's high, adaptive thinking always on, and a 400 error for a forced tool_choice.
• Rename the OpenAI guides from gpt-56-* to openai-gpt-*, target GPT-6 while keeping the GPT-5.6 guidance, and archive the superseded vendor docs verbatim.

Effort and refusals (work-orchestrator 4.5.0)

• State effort defaults per model, so an Opus 5.5 definition's rung is not read as Opus 5's.
• Re-route a safeguard refusal from a claude worker instead of accepting it as a verdict.
• Note GPT-6 in the GPT-5.6 docs and correct the GPT-5.6 prices.

Routing recalibration (work-orchestrator 4.6.0)

• Re-rank every codex model's severity labels before triage instead of sol's alone. Six repeated broad reviews each of gpt-5.6-sol, gpt-6-sol, and gpt-5.6-terra labeled a planted non-blocking defect as blocking in 17 of 18 runs.
• Drop the sol exclusion from security-flavored review. Four security reviews, sol included, met no safeguard friction.
• Limit the optional luna pass to adding a third model's coverage.
• Extend refusal re-routing to codex workers: a different codex model, a claude definition, or the session.
• State the sample size of every capability statement in docs/, record each routing row's basis there, and phrase per-model behavior in references/ and SKILL.md as working defaults. The measurements come from one small Python fixture, run through a proxy whose routing to the named models is unverified.