docs(agents): state the app-vs-platform boundary once, so it stops being re-derived

Three seats decided this ad hoc on one day and each derivation differed, because nothing
in AGENTS.md, CLAUDE.md or skills/ stated the rule. One new section, no other edit:

- the deciding question — could this be written by something that has only the metadata,
  and no knowledge of this company?
- the second question for publication — would a second app copy the implementation?
  (one consumer is a use, two is a contract)
- the two anti-patterns — an app hand-copying a platform rule, and a platform capability
  that under-delivers silently

Every lesson is carried self-contained (failure mode, discipline, boundary) with no
issue-ID citation, per the 2026-08-12 ruling that check:pm-skill-id-lint enforces.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UHvF5hyiZjnCyExFnfQB8m

Author: claude

AGENTS.md

@@ -962,6 +962,38 @@ Boot the real composition with its real services, or do not claim an answer.
 
 ---
 
+## App or platform — which work belongs where
+
+**The deciding question: could this be written by something that has only the metadata,
+and no knowledge of this company?**
+
+| Answer | Belongs in |
+|:--|:--|
+| **No** — it encodes the company's own judgement | the **metadata app**: a discount ceiling, who a case is assigned to, how won/lost is booked; its own objects, views and flows |
+| **Yes** — it only asks whether the metadata is self-consistent | the **platform**: reference integrity, translation coverage, view rosters, sharing-rule coverage, CRUD round-trips, RLS probes per declared position |
+| the subject is platform behaviour, the cost lands on the app | the **platform**, and it is a **gap** until it does — asserting what a hook does inside the platform's own sandbox is one |
+
+**Second question, for a capability an app wants published: would a second app copy the
+implementation?** Yes ⇒ platform. A resolution wanted by four call sites inside one package
+stays package-private on exactly that test — **one consumer is a use, two is a contract**,
+and a symbol never published still moves without a major.
+
+⛔ **An app never hand-copies a platform rule.** The copy diverges from the rule the build
+actually applies, so the app stays green while the platform refuses. A card wanting the
+platform's hook-body lowering pass turned out to want half a command that already ships
+(`os build --strict-body` fails a bundled-instead-of-lowered body with exit 1) and half an
+`os lint` rule that exists nowhere in the tree. One level up, the same shape ships a rule
+claiming "0 findings over the corpus" against a corpus that is not the app it names.
+
+⛔ **A capability that under-delivers silently is worse than none.** `os verify` derived
+**zero** cases on a multi-package app and reported success; `verify.ts` names the shape —
+*a verifier that under-verifies reports success it never established*. An app that meets one
+writes its own and never comes back, so the order is fixed: **make the derived half
+trustworthy first, then take the hand-written half back.** That fix landed as an ordered
+pair — ledger the losses, then close them; reversing it just deletes protection.
+
+---
+
 ## Degradation log levels — `warn` vs `error`
 
 Nearly every `catch` in this repo is a best-effort degradation, and nearly every one of