README: make the collected-data list match what is sent

[mattpodwysocki]

Follow-up to #17. Three mismatches between the Privacy section's collected-data list and telemetry::telemetry_markers, which is the whole of what goes out:

let mut markers = vec![os_marker(), arch_marker()];
markers.extend(ci_marker());
markers.extend(agent_detect::detect_agent().map(|a| format!("agent/{a}")));
markers.push(terminal_marker());
if let Some(group) = command_group { markers.push(format!("command/{group}")); }

Terminal state was collected and not disclosed — the direction that matters. terminal_marker() sends stdin_tty/… stdout_tty/… on every request; the list didn't mention it. The docs.mapbox.com page does, so the two disclosures disagreed and this was the one that under-stated.

Exit codes aren't collected. The User-Agent is assembled before a command runs, so there's no exit code in existence to send — nothing in telemetry.rs or http.rs reads one.

It's the service, not the command name. The marker is set in exactly one place — http::client_for(Some(op.service.as_str())) in executor::dispatch — so the value is styles or geocoder, never styles list. The old example was auth login, which is the one thing that can't appear: auth is hand-written rather than generated, never reaches dispatch, and uses client_for(None).

The new phrasing keeps that distinction rather than just deleting the wrong words, because the distinction is the reassuring part — which service was used, never the operation or its arguments.

Scope

Only that one sentence changed. Retention, third-party access, what we don't collect and the opt-out are all byte-identical — I checked by splitting the section into sentences with whitespace collapsed and diffing the sets, rather than by eye:

sentences removed: 1 (the What We Collect bullet)
sentences added:   1 (its replacement)

Some lines rewrap around the edit; no other wording moves. That was deliberate: the surrounding text reads like reviewed copy, so this sticks to claims that can be checked against the code.

Worth considering separately

Nothing stops this drifting again — the list is prose and the markers are code. A blunt guard in the spirit of tests/source_guards.rs would catch it: assert the number of markers telemetry_markers can emit, so adding one fails a test that names the README as the thing to update. Happy to add it if you want; it's a different kind of change from this one.