docs(benchmark): SWE-bench results at matched-50 + home-page rework

[OsherElhadad]

Summary

Updates the live SWE-bench Verified benchmark story to the matched-50 set and reworks the docs home page so the numbers are legible and stated as percentages.

The two tasks that previously hit infrastructure exceptions in the context-guru arm (sympy__sympy-23413, django__django-14792) were rerun — both now score cleanly and both solve. All three arms (baseline / context-guru / headroom) now cover all 50 tasks with zero exceptions, so the comparison set grows 48 → 50 and every number is apples-to-apples on the same tasks.

Headline (matched-50, live · Claude Code · aws/claude-sonnet-5)

dimension	baseline	context-guru	headroom
tasks solved	86%	88%	80%
billed cost vs baseline	—	−13.2%	−5.3%
cache-read tokens vs baseline	—	−17.8%	−6.3%
cache-write tokens vs baseline	—	−0.4%	−0.9%
mean steps / task vs baseline	—	−13.9%	−2.8%

• context-guru is the cheapest arm and now solves the most tasks of the three.
• Cache-write is a three-way wash (all within ~1% of baseline) — none of the arms busts the cache. The earlier "context-guru wins cache-write" claim is corrected to a tie.
• headroom keeps its one edge: lower added latency (63 vs 117 ms/req), since it runs no model on the hot path.

Changes

• Rows / analysis: merged the clean reruns into the canonical context-guru rows; regenerated the three-way deep analysis and the 6 benchmark plots at matched-50.
• Docs: context-guru.md (50 tasks / 44 solved), comparison.md (headline + verdict + honest caveat), RESULTS.md, components.md head-to-head, README.md headline, and charts.js inline chart data — all now percentage-first and consistent.
• Home page: reworked the "The numbers" section into a clear bottom-line conclusion + percentage stat cards with comparison context; bumped several too-small font sizes in the stylesheet.

mkdocs build --clean is warning-free; the full three-way comparison, per-task tables, and reproduce guide are all in the Benchmarks nav.