docs(readme): fix stale TTFT claim, clarify reuse-ceiling framing

[konjoinfinity]

Summary

Syncs the README's benchmark table against the now-corrected numbers in the blog post (docs/blog/posts/local-llm-fast-enough.md) and docs/paper.md §4.4, which this session already fixed after the same stale claim was caught there.

Two stale claims found and fixed:

• "Cold short-prompt TTFT: 167ms vs 192ms, Squish loses." This came from a flawed methodology (the same fixed prompt sent five times), which was actually measuring Ollama's own cache hit on the repeat sends, not a genuinely cold comparison. Replaced with the corrected cold/unique-prompt floor benchmark: Squish leads 800ms vs 812ms at 75 tokens (1.15× faster) — Squish has no measured TTFT loss anywhere under real cold, unique-prompt conditions.
• The same claim repeated in "What Squish Doesn't Do" as a listed limitation ("Slower first token on a cold, short prompt than Ollama"). Removed entirely, since it's no longer true.

One clarification added (not a correction, but avoids reproducing the same misleading framing the blog post's original stale draft had):

• The "Full response @ 4000-token prompt" row shows the 9.8× reuse ceiling (an identical prompt resent exactly) without qualification. Added (repeated exactly) to the row label and a footnote pointing to the real unique-prompt floor (1.15–1.32×), matching how the blog post now presents both numbers side by side.

Also found, out of scope for this PR: BENCHMARKS.md has the same stale claim in three different variants (lines 24, 70, 74 — including a third, different pair of numbers: 131ms vs 279ms). Flagging this as a separate follow-up needed since the ask here was specifically the README.

Type of change

• Bug fix
• New feature
• Performance improvement
• Documentation update
• Refactor / cleanup

Checklist

• No hardcoded absolute paths
• No model weights, eval output files, or log files staged for commit
• Changes are scoped to one logical concern (README benchmark accuracy)
• pytest tests/ passes locally — N/A, docs-only change
• ruff check squish/ tests/ reports no errors — N/A, no Python changed
• Performance-sensitive changes include a before/after squish bench run — N/A

Related issues

Follow-up to #215–#222 (the blog post sync/correction work this stale claim was originally caught in).

---

Generated by Claude Code