BigMoeOnEdge

Adds lfm2moe to the architecture registry: Liquid AI's LFM2 and LFM2.5 MoE models
(LFM2.5-8B-A1B, LFM2-24B-A2B) now load and stream. One registry row, no engine change.

Why it is one row

LFM2 MoE is a hybrid short-convolution/attention stack, but its MoE blocks name their experts with
the standard split layout and route through the same helper every supported MoE uses, so the routing
node the engine hooks is the one it already reads. Expert count and per-expert stride are discovered
from the model file at runtime, as for every other architecture.

Two structural properties are documented rather than left to be rediscovered: the leading dense
blocks name no expert tensors and stay resident, and the router's per-expert bias applies before the
top-k, so the node the engine observes is unaffected.

Scope

Registering an architecture states that the layout streams losslessly. It does not claim streaming is
the right way to run these models: at ~8B total, LFM2.5-8B-A1B fits in phone RAM, and a model that
fits is faster loaded resident. docs/limitations.md now says so explicitly. The model is
deliberately not in the in-app download catalog, which lists the models the engine is measured on —
use the URL field or the file picker.

Verification

Byte-identity gates pass; the split-layout gate already covers this streaming path. Validated
on-device against LFM2.5-8B-A1B (Q5_K_M): experts stream with O_DIRECT over the discovered bank of 32.
No benchmark tables changed — this was a functional check, not a measured configuration.

Known issue

The Thinking toggle has no effect on LFM2.5: its chat template implements no off switch, so the
request is silently ignored and the model reasons anyway. Tracked in #82 with the fix direction.

Install: app-dev-release.apk.