"Mens LoRA / adapter ownership (vox-tensor vs vox-populi)"

Mens LoRA / adapter ownership (vox-tensor vs vox-populi)

Split

Crate / tree	Owns	Do not duplicate here
`vox-tensor` `crates/vox-tensor/src/lora.rs`	Low-level LoRA linear math, parameter layout, and shared tensor utilities consumed by graph code.	HF-specific key maps, QLoRA export, merge-CLI, or `training_manifest` fields.
`vox-populi` `crates/vox-populi/src/mens/tensor/lora.rs` + `lora_vox_transformer.rs`	Transformer-shaped LoRA modules, Burn training graph, checkpoint (`.bin`), merge* for Burn, and integration with `FineTuneContract` / planner.	Re-implementing generic rank decomposition — call into `vox-tensor` where appropriate.
`vox-populi` `candle_qlora_*`, `qlora_preflight`, `adapter_schema_v3`	Candle + qlora-rs QLoRA train/export, v2/v3 adapter manifests, `merge-qlora`, HF shard/key inventory.	Burn `*.bin` merge path (`merge-weights`).

Any change to LoRA scaling (alpha/rank), merge equation, or adapter tensor naming must either touch one canonical implementation and call sites, or be documented as an intentional fork with a test linking both behaviors.
PRs touching both trees: use mens-llm-pr-checklist.md and add/adjust a regression test in the kernel that actually runs the changed path (cargo test -p vox-populi --features mens-train …; vox-tensor unit tests for primitives).