Module hidden_trace 

Env-gated hidden-state trace, shared by the reference executor and the CUDA trunk.

MEMRA_HYPER_TRACE=<path> makes BOTH sides append the same stage names, in the same format, from the same token ids — so a bisect compares like with like instead of two hand-rolled dumps that agree on nothing but the layer index. It lives in the reference crate because memra-engine already depends on it, so one emitter serves both.

Format (memra-hidden-trace-v1), one line per stage, the LAST token row only:

stage\t<name>\t<layer|-1>\t<width>\t<f32 bits hex>,<f32 bits hex>,...

Only the last row is emitted: it is the row the logits come from, and it is the row every banked oracle in this lane already pins. Width is streams * hidden at the residual-stream stages and hidden at the branch stages.

Off by default and, when off, costs one OnceLock read per call — no allocation, no device work, and no arm selection changes anywhere (unlike the MoE traces, which move observation_mode). Turning it on cannot change what the model computes.

Functions

emit_last_row

Emit the last token row of a [rows, width] row-major activation.

emit_tokens

Record the run’s token ids once, so a trace file can be checked against its oracle TSV.

enabled

True when MEMRA_HYPER_TRACE named a path. Callers use it to skip a device readback.