docs(v5): wide 25-gap epic — math-effect, distributed, lifecycle, persistence

Honest V4 audit found 25 remaining gaps in 9 groups. Most damning:
V4-7/V4-8/V4-10 are propagation-theatre tests — UI dropdown changes a
string, extras shows the same string, training math stays unchanged.

v5 closes string-echo→math-effect for all 11 such tests, fills sharding/
distributed dim (G05-G07), adds lifecycle (G08-G10 cancel/abort/reconnect/
warm-start), persistence (G11 save/load, G12 checkpoint resume), tightens
weak coverage (G13-G16), adds side-channel forward routing (G17), closes
UX feedback loops (G18-G19), real-tokens inference probe (G20), other-
stages observability (G21), optimizer math verification (G22-G23),
roundtrip byte-exact (G24), and MoE routing observability (G25).

Epic cppmega-mlx-pjt + 25 gap tickets (cppmega-mlx-iry/3tu/7w9/3m6/8be/
cn3/wz2/406/ajm/jrp/4n3/t8i/ay6/amy/a0b/3va/7ei/biy/udk/5kt/aih/3p2/br8/
edp/zuf). 195 sub-tasks total across 25 gaps.

VisualBuilderPlan-v5.md: 350+ lines, group-by-category breakdown,
per-gap acceptance + sub-task plan, workflow, acceptance counters,
bd-id mapping table.

VisualBuilderSpec-v5.md: 400+ lines, post-v5 extras schema (30+ keys),
per-gap backend kernel pseudocode, frontend testid contracts, file
layout for 18 new pytest + 25 new e2e specs.

Honest-categorisation rule: every UI mutation gets a label
(🟢 math-effect / 🟡 propagation / 🔴 decorative). Target: 0 🔴,
🟡 only for genuinely-config fields, ≥20 🟢.

Author: apstenku123