feat: add Claude Opus 5.5 as a benchmark pricing model

[Mapleeeeeeeeeee]

Summary

• Add --model opus-5-5 → token counting with claude-opus-5-5, pricing $4 base input / $5 cache write (5m) / $0.20 cache read.
• Opus 5.5 cache read is 5% of base input (official: "On Claude Opus 5.5, a cache hit costs 5% of the standard input price ($0.20 USD per million tokens)."), so the comment and docs that said "every non-Fable model uses 10%" are rewritten to list 10% / 5% / 2.5%.
• Default opus alias stays on opus-4-8 so existing baselines stay comparable.

Source: https://platform.claude.com/docs/en/about-claude/pricing (Opus 5.5 released 2026-09-22; 1M context at standard pricing, no long-context tier, so Pricing needs no new fields).

Verification

• go build ./..., go vet ./..., go test ./... pass
• cc-session benchmark --help lists opus-5-5
• Benchmark, baseline flags -n 20 -min-kb 100 -days 90 -overhead 40731, same 20 sessions, run 2026-09-23:

Metric	opus (4.8)	opus-5-5
Compression median	75.8%	75.7%
Cold cache: break-even / 10-turn / 100-turn	turn 1 / 65% / 35%	turn 1 / 64% / 35%
Warm cache: break-even / 10-turn / 100-turn	turn 1.5 / 57% / 35%	turn 2.5 / 49% / 33%

NewCtx differs by only 2 tokens per session (Opus 4.8 and 5.5 share the post-4.7 tokenizer). The warm-cache drop comes from the 5% cache-read ratio: staying in the original session re-reads its large cached context more cheaply, so the new-session advantage shrinks.

🤖 Generated with Claude Code

https://claude.ai/code/session_01FxSSNGDv4kKwZZ3oHdhbef