docs(adr): ADR-0087 v2 — redesign for AI consumers: conversion over notification, replayable cross-major migration chain

Two review corrections reshape the design:

1. The consumer's maintainer is an AI agent, so human attention is not the
   scarce resource — machine executability and verifiability are. Notification
   (CI warnings, prose guides) demoted; a load-time conversion layer (most
   breaks need zero consumer action), executable migrations, structured
   fail-fast diagnostics, and an MCP upgrade surface promoted. Framework-side
   smoke ring dropped in favor of consumer agents pulling RCs and verifying
   themselves. Includes the PD#12 boundary argument for why a centralized,
   versioned, expiring conversion table is not a consumer-side dialect
   fallback.

2. A one-major warning window is worthless to a consumer upgrading across
   multiple majors. Timeliness is now explicitly non-load-bearing: retired
   conversions graduate into a permanent, ordered, fixture-tested migration
   chain; 'migrate meta --from N' replays any past major to current in one
   command, with per-hop verify checkpoints; per-major change manifests
   compose into a single from->to view.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HjDghc79qFSJkJt2zvBxtU

Author: claude