feat(panel): warn low-memory Macs about the in-RAM cache, in every language

The engine already advises SSD-only caching at session startup, but that line
lives in the server log where most app users will never see it. The people who
most need the advice — 16 to 32 GB Macs, where the in-RAM paged KV cache
competes with model weights for the same unified memory — are the least likely
to be reading engine output.

The recommendation now appears at the top of the Prefix Cache section, above
everything else in it, styled as a warning rather than a hint, and only when
the RAM cache is actually on (there is nothing to recommend once it is off).
It quotes the machine's real memory and the measured cost: within about 2
percent of SSD-only on a 15k-token workload, so following it gives up almost
no speed while freeing gigabytes for weights.

Translated into all five shipped locales rather than English-only, because a
missing translation silently falls back and this is precisely the audience
that would miss it.

The wording states plainly that nothing has been disabled, and a test asserts
that property in every locale: an invented RAM guard once refused to load big
models across six releases, and this must never read as though the app acted
on the user's behalf. Detection failure stays silent for the same reason —
a guess that cannot be made must not become a warning.

Threshold mirrors the engine's own (36 GB), with a comment pointing at
_low_ram_cache_advice_lines so the two cannot drift apart unnoticed.