ENG-HYBRID-PLACEMENT: map llama.cpp's four placement surfaces onto --offload-config's vllm_cpp key, and re-anchor the spec at b10451

[localai-bot]

ENG-HYBRID-PLACEMENT is READY with a committed spec and no implementation. Two
things about that spec have to change before a wave builds against it, and the
developer has settled both (2026-08-26).

The surface moves onto the existing configuration. The spec proposes a bespoke
pattern-to-device flag. The direction is to map llama.cpp onto the vllm_cpp
extension key of --offload-config, which already carries mmap, expert_stream
and device_fit, and which already has a settled precedence rule of environment
variable over JSON over built-in default. One user-facing flag stays one user-facing
concept, and the mirrored OffloadConfig fields stay byte-faithful transcriptions.

All four llama.cpp surfaces are in scope, not two. -ot <regex>=<buffer type>
is the general form. -cmoe and -ncmoe N are literally sugar over it: at the pin
each one pushes entries into the same tensor_buft_overrides vector. --fit is the
auto-resolver, and it writes into that same vector and refuses when the user already
set one. Building the general form first and deriving the specific ones from it is
llama.cpp's own shape, and it is the shape that composes with ENG-WEIGHT-OFFLOAD
and #147 later.

The spec's anchors are stale. It cites 237ad9b96 (b9892). The recorded pin in
.agents/oracles/llama-cpp.md advanced to b10451
on 2026-08-22, and the cited line numbers do not resolve there: common/arg.cpp:2451-2478
is now :2715, :2721 and :2728, and common/common.h:1046-1054 is now :1113-1120.
Re-anchoring is part of this change.

Scope is the spec, the row's matrix cells and its claim. No product code. The waves
this spec then dispatches get their own issues.

Owning row: ENG-HYBRID-PLACEMENT. Campaign issue: #149, which stays open because it
also covers the dense half (ENG-WEIGHT-OFFLOAD) and the multi-GPU half (#147).