docs: align README benchmark claims with the agentic results

[ggeandersonreis]

Docs only: no ruleset, skill or hook changes. Every number here comes from benchmarks/results/2026-06-18-agentic.md; nothing was re-measured.

What changed

• "ponytail is the only arm that cuts every metric" (README and results file). The README's own table shows the "YAGNI + one-liners" prompt also cuts all four (−33% LOC, −14% tokens, −21% cost, −30% time), and it is slightly cheaper and faster than ponytail. The sentence now states what does set ponytail apart: by far the biggest code cut, and the only arm that cuts every metric while staying fully safe (the YAGNI prompt dropped a path-traversal guard in 1 of 20 runs).
• Chart alt text: "lowest on every metric" → lowest on LOC and tokens; the YAGNI prompt is slightly lower on cost and time (79% / 70% vs 80% / 73%).
• Headline: "100% safe" → "safe in 20/20 adversarial runs", and "up to 94%" → "up to 94% on one task" (the date picker). Without the qualifier, the 94% reads like the single-shot 80–94% figure the README itself retracts further down.
• Codex install: "trust its two lifecycle hooks" → "trust its lifecycle hooks". .codex-plugin/plugin.json points at hooks/claude-codex-hooks.json, which defines three (SessionStart, SubagentStart, UserPromptSubmit). If Codex really shows two, putting the number back with a note on why would be clearer than my edit.
• Hosts with commands: added Kimi Code and Goose, which INSTALL.md documents as exposing the six commands.

Not changed, for you to decide

• Translations: i18n/README.{es,ko,zh-CN,ja}.md carry the same headline, alt text and host list. I left them alone rather than hand-edit Korean, Chinese and Japanese prose.
• Results file: the safety section's header and summary say "6 safety tasks", but the safe-rate table says "5 security tasks × 4 runs = 20". It doesn't say what the sixth task is.

🤖 Generated with Claude Code

[DietrichGebert]

Thanks! Ponytail 5 replaced the headline, the chart text and the "only arm" sentence with the new benchmark, so these edits don't apply anymore.