docs(plan): resumable execution plan for the projection/payload vocab batch

Sequences the four remaining linked units — #270, the doctrine docs, the @ROLE
shrink (subsuming #212), and #210 — into two coordinated releases, written so a
cold session can execute it from the file alone: exact paths, verified line
anchors, verbatim gate commands, and per-unit stop-and-escalate signals.

Two dependencies the sequencing turns on: #210 is blocked on #270 (the payload
tier must be declared-authoritative before values lose assembly origins) and on
#271's disposition, which the roadmap ruling names as a second prelude. #212 is
subsumed by the @ROLE shrink — its actionable registry content is exactly
"remove publish".

Release shape is two cuts, not one. #270 is a live silent-wrong-output bug (a
declared curated payload VO becomes the full entity in Kotlin/Python output) and
ships as its own coordinated PATCH rather than waiting on the breaking batch.
The @ROLE shrink and #210 are author-facing vocabulary removals and ride one
coordinated MINOR — the ADR-0035 §3 consolidation window #210's ruling requires.

Gate 0 for the @ROLE shrink is recorded as CLEARED: the ADR-0036 adopter scan
found zero uses of index/cache/publish/mirror anywhere — this repo, the public
reference app, and downstream consumer models on the workstation. The only
@ROLE values authored anywhere are primary and replica.

The plan opens with the meta-lesson that made it necessary: two issues filed
2026-08-05 both had their premises falsified by recon within hours, so every
unit begins with a premise-recon step and instructions to stop and report rather
than force a falsified plan through.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S3msoGxjRMwx94PhKSLDuE

Author: dmealing