0.70.2 P0: ladder fit gate (#4016) refuses every .apr model (llama-fit-params cannot read .apr; exit 1 is read as does-not-fit) #4668

What fails

On lambda (0.70.1, cc4463f), the ladder refuses all five .apr inventory rows before it calls apr at all:

[REFUSE] inv:qwen2.5-coder-0.5b-instruct.apr fit: does-not-fit -- llama-fit-params exit 1: the model does not fit this host's free memory

The same happens to qwen2.5-coder-1.5b-instruct-q4k-v2.apr, -q4k.apr, -st.apr and qwen2.5-coder-1.5b-q4k.apr. On gx10, three .apr rows are refused the same way.

The files are 1–3.5 GB, and the GPU has about 24 GB free (on gx10, 128 GB unified memory). llama-fit-params is llama.cpp's tool: it does not read apr's format and exits 1. The fit gate (#4016) reads that exit as "does not fit". check_model_ladder.sh then counts each row as FAIL: no qa_rc, and no "fits" verdict.

v0.69.5-rc.1 is refused the same way, because the fit gate runs before the binary is called. Under ruling C272 5b / C276, 0.70.1 ships these rows as known red, for 0.70.1 only.

Done when (0.70.2)

• The fit gate gives .apr files a real fit verdict, or refuses them with a reason that says what actually happened. It never reports a tool error as "does not fit".
• The 0.70.1 known_red entries for these five files are removed. Every .apr row is measured, or refused with its true reason.