docs: two prose claims still assert alias parity as current fact

[ric03uec]

Split out of #1. That issue's four Impact items all landed; this is the residual
prose that still states alias parity as a fact rather than an aspiration.

tests/parity.yml is now empty on purpose and its comment block explains why.
Two files were not updated to match.

Tasks

1. references/model-catalog.md, the "Aliases and parity" section — the
   Parity bullet says qwen2.5-coder-7b "means the same thing on the NVIDIA
   host and the AMD host". It does not: h2-amd serves hermes-3-llama-3.1-8b.
   Reword the bullet to describe what virtual_models is for while stating
   that no alias is currently required in common, and that tests/parity.yml
   enforces the list only when it is non-empty.

2. AGENTS.md, invariant 3 "Alias parity" — reword so it does not assert parity
   as a current invariant. It is a goal that tests/parity.yml enforces only
   for hosts listed there, and that list is empty.

Pattern to follow

tests/parity.yml's header comment is the precedent for the wording and the
level of detail. Mirror its framing.

Out of scope

Do not touch README.md, website/, or PLAN.md. Those carry the same claim
inside ASCII diagrams and are being rewritten separately.