docs(ci): haus's acquire holds two suites, and a step's floor is not the pole

[JulienMartel]

What

Two edits to docs/ci.md.

1. The haus row said the nix half is "nix eval …, nix flake check alone, haus add alone". acquire carries test/haus-settings.sh as well since ci: test/haus-settings.sh runs on the Linux runner, and always could haus#766 — the other suite that drives a real consumer flake through a real nix, and the last bash suite in that repo CI was not running at all.
2. Rule 5 gains a fourth worked example.

Why

The row was simply stale. The rule-5 addition is the part worth reading.

Rule 5's three existing worked examples — haus's agents (no), perch and trill (yes), scruff (neither) — all answer the same question: does this job divide? hausfold/haus#766 asked a different one and the rule had no written answer for it: is this step worth a job of its own?

haus's nix half had a local heuristic for that — a new step is measured against the ~35s cache restore standing in front of it, and above that bar it earns a job. test/haus-settings.sh came in at 66s, twice the bar, and joined an existing job anyway, because the pole in that gate is nix flake check at 243s and a fourth job cannot take a second off it. The job went 46s → ~112s and the run cost what it cost before.

So: a fixed cost says what a job is worth paying for. It says nothing about what a step is worth moving, and reading it as a floor sends you to a job that buys the gate nothing. The gap to the pole answers both.

Verify

• docs/ci.md only. No code, no workflow, nothing bench reads.
• The haus side is ci: test/haus-settings.sh runs on the Linux runner, and always could haus#766, and the figures quoted here come off its runs: the suite's own 35594772634 (66s cold, 51s warm) and the job table on 35594402514 (pole checks 243s, runner-up eval 75s, acquire 46s).

Watch out

• This lands a claim about haus's CI shape, so it wants ci: test/haus-settings.sh runs on the Linux runner, and always could haus#766 merged first or the row is briefly ahead of the repo it describes.
• script/probes/README.md's stamped figures are untouched and still correct — #766 changes no job those measure.