docs(benchmark): add rtk (Rust Token Killer) as a 4th SWE-bench arm

[OsherElhadad]

Summary

Adds rtk (rtk-ai/rtk, "Rust Token Killer") as a 4th arm to the live SWE-bench Verified study, alongside baseline, context-guru, and headroom.

rtk is architecturally distinct from the two request-stream proxies: it is a Claude Code PreToolUse hook that rewrites Bash commands (cat→rtk read, pytest→rtk pytest, grep→rtk grep, …) and compresses their output at the shell, before it enters the transcript. It is not on the network path — so it adds 0 ms request latency, runs no model ($0 tool cost), and is cache-safe by construction (the compressed form is what gets cached; there is no cached prefix to invalidate).

Result (matched 50 tasks, all four arms, 0 exceptions)

dimension	baseline	context-guru	headroom	rtk
reward (solved / 50)	43	44	40	43
billed cost	$31.98	$27.77 (−13.2%)	$30.30 (−5.3%)	$29.09 (−9.0%)
cache-read	102.8M	84.5M	96.4M	91.7M
mean steps	36.1	31.1	35.1	33.2
added latency / req	—	117 ms	63 ms	0 ms
tool LLM cost	—	$0.31	$0	$0

rtk is the efficiency surprise — a simple deterministic shell filter is the 2nd-cheapest arm (−9.0%), reward-neutral vs baseline (43 = 43), at zero request-path latency and $0 tool cost, beating the headroom proxy on both cost and reward. Its 65.8% bash-output reduction (dominated by cat→read and grep) nets to −9% on the bill because bash output is a small slice of a ~98%-cached agent's context; its ceiling is that only Bash-tool output passes the hook (Claude Code's built-in Read/Grep/Glob bypass it), which is why context-guru's whole-request compaction goes deeper.

Changes

• New docs/results/rtk.md — per-task table, rtk's own savings ledger, per-command breakdown, and real before→after examples.
• Four-way updates to comparison.md, components.md, RESULTS.md, README.md, the home page, charts.js, and the mkdocs nav.
• REPRODUCE.md documents the rtk arm end-to-end.
• Harness: deploy/harbor/swebench_rtk.py + a committed reference copy of the custom claude-code-rtk Harbor agent; deep_analysis.py / deep_plots.py extended to a 4th arm (figures regenerated).
• mkdocs build --clean is warning-free.

Notes

• Model routing for the rtk arm is identical to the baseline (same off passthrough proxy) — the only variable is the in-container bash compression.
• The content-removed metric is not comparable across arm kinds: the proxies measure against the whole request; rtk's 65.8% is of bash output only.

Assisted-By: Claude (Opus 4.8)