fix(engine): stop making the unused vision encoder resident (852 MB, every tier)

[penta2himajin]

The checkpoint ships a vision encoder. qwisp is a text-only engine and never reads one of those tensors — but WeightStore.residentNonExperts() filtered only on .switch_mlp., so all of vision_tower.* went to MLX.eval() and was made resident on every run, on every tier. 333 tensors, 852 MB.

Measured

Byte-exact against the prediction made before running it:

tier	before	after
8GB (budgetC=64)	7,071 MB	6,219 MB
16GB (budgetC=128)	11,391 MB	10,539 MB

Which confirms the decomposition those predictions came from:

active = arena(40 x budgetC x 1.6875 MB) + language_model non-expert 1,325 + scratch 574

Output unchanged (it was never read).

Filtered at load, not at eval

Keeping the MLXArray alive holds its mmap slice, so removing it from the eval list alone would not have returned the bytes. residentAll() becomes correct for free — the tensors are simply not in the store.

Why this matters beyond the byte count

• 8GB tier. It needs ~761 MB to hold coverage at or above the ~103-expert per-layer footprint (the LOOPY-safety line). 852 MB is more than the entire deficit, at zero quality cost. This session had been measuring whether non-expert weights survive 2-bit in order to find those bytes. They do not — GDN loses 31 points of teacher-forced agreement below its own same-bit round-trip floor (94.3% → 62.9%), lm_head 24.6, embed 6.3 — and quantizing every non-expert tensor would only have freed 589 MB anyway. The bytes were already there, being wasted.
• 16GB tier. The bolt: decode command buffer fails with GPU OOM and cb.error is never checked → unbounded token-id-0 (!) output above ~2–6K prompt tokens #169 OOM cliff was measured at two experts per layer, ~150 MB (coverage 228 passes, 230 fails). 852 MB is 5.7x that margin.

Not claimed here

Whether this makes the 8GB tier actually fit a real 8GB Mac, and whether it moves the #169 threshold, are follow-up measurements — both need the wired-limit emulation and are not in this PR. This PR is the byte reclamation and its proof.

Gates

RAWTESTS 100/100 · BENCHBATCHTEST PASS · COMPTEST 97/97 · CBGUARD PASS

The new locked test asserts the predicate keeps engine tensors (embed, lm_head, switch_mlp, linear_attn, mlp.gate) and drops vision ones — a regression would silently cost 852 MB again. Runs without a model or GPU.

🤖 Generated with Claude Code