Concurrency and scaling

Scope: this implementation’s own serializers and their measured cost on one many-core host. Not normative. Target: the C++23 reference implementation under core/, on a many-core host. Companion: ../../reference/15-concurrency-and-scaling.md is the standard-level half — the obligations any implementation must meet and the four hardware regimes. Everything here is specific to this codebase and one host, and none of it is normative.

Where the reference suite states what must hold, these pages state what it costs here: which serializers exist, what each one guards, which shapes scale on this host, and which serializers remain on the hot paths.

File	Topic
00-scaling-and-serialization.md	The serializer inventory with file anchors; the two independent limits on the read path; the measured effect of removing the per-read map lock and of binding the hazard-pointer slot; the cost budget; the remaining serializers; the diagnostic recipe; and how measurement goes wrong here.
01-write-and-delivery-path.md	The write and delivery half: the snapshot-under-lock split, the vertex lock stripes taken on every delivery, fan-out and scatter-gather delivery cost, and the forward-demux mount scan. Its forward-hop and mount-scan figures come from bench/bench_forward_demux.cpp, the instrument that settled the acceptance condition of ADR-0061, per-module mount routing.

The three instruments behind the figures here

bench/ holds several dozen bench sources. This table is not an inventory of them: it names the three that produce the numbers in 00-scaling-and-serialization.md, because it matters which one a figure came from. The write and delivery path quotes a fourth, named in its row above.

instrument	what it measures	committed?
bench/bench_contention	the machine — what one shared cache line costs, libtracer absent	yes
bench/bench_lkv_slot	model slot policies (slot mode) and the real graph_t::read/write shapes (graph mode)	yes
the has_registered_child ablation	the ceiling a lock removal could reach	not committed; it short-circuits the fork check (vertex_t::has_registered_child, core/include/libtracer/vertex.hpp:916, called from core/src/graph.cpp:1837), which breaks the composed branch read

A slot-mode arm models a slot policy in isolation and is not the cost of any code path. The arms are named model_* in the source (bench/bench_lkv_slot.cpp:193,342,456) so that a model policy and a real library type cannot be confused. Never quote a slot-mode number as a forecast for a real path.