fix(minimax-h3): the fp32-island rule now BINDS the bf16 GGUF host loader (#244)

[localai-bot]

Closes #244.

What #253 fixed, and what it left

#253 fixed the instance — rope.inv_freq and adaln_t_table stopped being rounded into bf16 bits that Ptr<float>() then reinterprets as garbage. But it did so by hardcoding those two names. MiniMaxH3IsFp32IslandTensor covers seven: both patch projections, time_embedder.*, both output heads, and those two. Five islands were still silently down-converted on this path.

Severity, stated precisely

The issue reads worse than it is, so: those five are not garbage. BindMiniMaxH3DitViews types each tensor from the storage map it lands in, so a bf16 island is correctly typed bf16 and consumed as such. rope.inv_freq was uniquely dangerous because the forward reads it through a raw Ptr<float>(), bypassing the dtype.

What remained is a silent precision divergence from upstream on one loader, not corruption — and on the CPU --dequant-bf16 path only. The production GPU stream was never affected, which is why every render this session was correct.

The fix

Consult the single source instead of re-listing names. The rule's own contract says "all four staging paths must agree" — this host loader is a fifth that no gate covered, which is exactly how the two descriptions drifted apart.

The gate is the load-bearing half

It walks every tensor the loader placed and requires island ⇒ storage (f32), non-island ⇒ bf16_storage, with an islands >= 6 floor so a predicate matching nothing can't pass vacuously. It asserts on the loader's own storage split rather than the bound views, because that map is the decision under test.

Mutation-verified

	test cases	assertions
with the fix	79/79 pass	57,395 / 0 failed
reverted to the hardcoded two names	78/79 — 1 failed	24 failed

The 24 are exactly the five island categories the hardcoded check missed. Tree restored byte-for-byte afterwards.

Suite grew 57,299 → 57,395 (the gate adds 96 assertions), so it is genuinely executing rather than passing vacuously.

Pattern

Third instance today of a rule documented as universal that one code path quietly ignores — after decoder_tiling "defaults false" (#251) and vllm_server_main compiled out under VLLM_CPP_SERVER=OFF (#202). Cheap to fix, expensive to rediscover.