cachesplit's -34.1% result is described as general, but it only applies to Claude Code CLI traffic

What's wrong

config/config.go:369 describes cachesplit as:

It is also the best-evidenced single component in the repo: -34.1% cost and 0% ->
96.7% prefix-cache hit in an isolated A/B (docs/results/context-guru.md).

Three things about that are not right, and the repo already contains the corrections.

1. The cited doc is not an isolated cachesplit A/B. docs/results/context-guru.md
line 1 is titled "context-guru codesmart (final) (SWE-bench Verified, 50 tasks)" — a
whole-pipeline run. Nothing in it isolates the volatile-tail split.

2. The figure is workload- and CLIENT-specific, and two other docs say so.

docs/results/terminal-bench-comparison.md:155:

cachesplit acted=0 for a structural reason: TB runs the Claude Agent SDK, which
never appends the git/env snapshot the CLI does, so all 73 captured requests carry 3
system blocks and zero volatile-tail markers. The -34.1% split figure came from
SWE-bench CLI traffic and does not transfer.

docs/results/improvement-plan.md:378 records the same thing as "zero legal opportunity".

3. It sits next to a much smaller production number. docs/results/measured-2026-08.md
credits the prefix split with $0.03 over the window (line 18), and line 179 explains it:
"1,799 mutations, credit granted on 3 requests". proxy/promexport.go:878 says to expect
it small on a warm-cache workload and calls the metric "A floor" — because the CLI
captures its environment block once per session, and 1,105 of 1,127 session starts were cold.

Why it matters

A reader deciding whether to run the component — or writing docs downstream of this comment —
takes "best-evidenced single component" and "-34.1%" as a property of the component. It is a
property of one workload on one client. The plugin's README and its plugin.json both lead
with the cache preset, so this comment is upstream of user-facing copy.

This is the defect shape already recorded twice on #249/#250: a measurement taken on one
workload, restated as a general property, surviving in every file nobody thought to grep.

Suggested fix

Documentation only — no behaviour change, and no argument here that the component should be
tuned or removed.

• In config/config.go, name the workload and the client in the claim, drop "isolated A/B",
  and point at the two docs that bound it.
• State the production figure alongside it, so the comment carries both.
• Check whether the same sentence has been copied into docs/reference/presets.md or the
  plugin's plugin.json description.

What is NOT claimed here

The mechanism is sound where the shape exists: it removes no content, is byte-identical in
concatenation, fails open, and is a documented no-op on implicit-cache backends. This is
about the evidence sentence, not the code.

Start with the cachesplit comment at config/config.go:369, then compare the cited evidence in docs/results/context-guru.md, docs/results/terminal-bench-comparison.md, docs/results/improvement-plan.md, and docs/results/measured-2026-08.md. Check docs/reference/presets.md and the plugin's plugin.json for copied wording; done means the documentation scopes the result to Claude Code CLI and the relevant workload, includes the production figure, and avoids implying a general component property.