docs(wiki): phase B training path and hardware decisions — recorded, not built

Claude-Session: https://claude.ai/code/session_01FWncR69HgGPuei1dLrW3cU

wiki/log.md

@@ -3133,3 +3133,13 @@ in-process) + `enqueueEntry()` (crop + camera class, no order/operator/category)
 kind column, operator agreement is wash-only. Setup line shows "1 in N entries sampled". Also:
 Setup → Car wash is a two-column grid (the master-data card was squeezed at max-w-2xl). Tests
 on both sides. Updated [[vision-review-outbox]].
+
+## [2026-09-07] decide | Phase B training path + hardware — recorded, not built
+User asked "now what about the training" and then "let's talk hardware". Recorded on the new
+[[bodytype-classifier-training]]: the five-step loop (train on the collector host → evaluate with a
+floor → publish weights to the registry → bake into the vision image → TAG bump; a booth gets a
+model the way it gets code, never a runtime fetch); ~200 reviewed crops per class before the first
+run; the Quadro FX 3800 is unusable (cc 1.3), the HD P530 irrelevant, the Xeon E3-1225 v5 is enough
+(feature-extraction head in minutes, full fine-tune ~1 h); trainer image = CPU-only torch, the
+compose seam drops the GPU reservation; cloud GPU rejected (crops stay on premises). Linked from
+[[opencv-anpr-service]], [[vision-review-outbox]], index. User: "No build just yet."