docs+fix: le modèle de sécurité local-first, et le retrait de linkTo qui le brisait

La règle d'accès est « qui a la référence ET la clé lit » — jamais « qui a la
référence lit ». Il n'y a pas de tiers dans cette phrase : le moteur ne vérifie
une permission qu'à l'ÉCRITURE, jamais à la lecture. Le contenu est chiffré, la
clé EST le droit. C'est ce que veut dire local-first ici : il n'y a personne à
qui demander, donc la possession est tout le mécanisme.

Ce que cette règle achète, et qui n'était consigné nulle part : une référence
n'accordant rien, **elle n'est pas récursive**. Un document largement diffusé
peut pointer vers un document restreint — la référence dit qu'il existe, la clé
dit qui le lit. D'où la confidentialité COMPOSABLE : un sommaire diffusé qui
renvoie à des chapitres restreints, un événement public qui renvoie à sa liste
de participants. L'auteur diffuse un document et décide encore, document
référencé par document référencé, qui reçoit la clé.

`linkTo` détruisait cette construction. Il rendait la CLÉ là où un appelant
demande une référence, transformant la règle en « qui a la référence lit » —
pour ce document et pour tout ce qu'il MENTIONNE. Ajouté et retiré le même jour.

Documenté en §0, avant tout le reste, avec la raison pour laquelle un agent
perd ce point avec constance : les réflexes client-serveur fournissent la moitié
manquante sans qu'on s'en aperçoive — quelque part un serveur vérifierait, un
lien serait un identifiant inoffensif, « Bob a-t-il le droit ? » aurait une
réponse. Aucun des trois n'est vrai ici, et du code écrit là-dessus n'échoue
pas : il accorde en silence. Rien ne passe au rouge.

Renommé au passage : `publishRepoLink`/`isPublished` → `recordInPublicStore`/
`isInPublicStore`. Ces méthodes n'ont pas de pendant amont et « publier » ne
désigne rien de précis ici.

171 tests unitaires, e2e 42/42 en 3,6 min.

docs/readcap-and-nuri-model.md

@@ -16,12 +16,47 @@ Purpose: to give the ground truth of NextGraph's access-rights model, in order t
 >
 > - **A comment describing the CURRENT state is not the intent.** §3's DIRECTION block exists because `RepoLinkV0`'s comment was read as the target model. It is not.
 > - **A word you recognise probably does not mean what you think.** `branch` is not git's. `wallet` is only a keyring — what we call a virtual user is a **user** (a *site*). Check the type before using the word.
+> - **The access rule is "whoever has the reference AND the key reads" — never "whoever has the reference reads".** An agent forgets this with remarkable consistency, because client-server reflexes supply a missing half: somewhere a server checks. Here nobody checks (§0 below). Getting it wrong does not fail — it silently grants. Every time you hand out a value, ask which of the two you just handed.
 > - **NextGraph's naming ALWAYS wins over any rule of ours — including the one that follows.** Where upstream has a word for something, that is the word, full stop. `publisher` is a case in point: it appears 126 times in the engine (`as_publisher`, `publisher_advert`) for a **pub/sub role on a topic**, and it must be used, unchanged, whenever that is what is meant.
 > - **What is banned is OUR loose "publish a document"**, which has no upstream referent and covers three different acts: *placing a document in a public store*, *making it findable*, and (in this emulation only) *handing out a key*. Every design discussion that used it drifted between them. NextGraph knows only the first — a document IS IN a public store, and brokers serve it accordingly. Say which act you mean. If you catch yourself writing "publish a document", you have not yet decided which one. This never licenses renaming something upstream calls `publisher`.
 > - **"I looked and it is not there" is not a finding.** §4quinquies once stated that no register existed for received caps, after checking one code path. `AddLink` had been sitting next to `AddRepo` in the same file the whole time. Absence needs at least as much evidence as presence — and an implementation *cache* (like local user storage) is never the model: it is what the model fills.
 
 ---
 
+## 0. The security model, and why an agent keeps losing it
+
+**This is the one thing to hold on to.** Every design mistake this file records — a cap minted where a reference sufficed, a link that granted what it was meant to merely name, an "is it allowed?" call that cannot exist — comes from letting go of it.
+
+### The rule
+
+> A document is readable by whoever holds **the reference AND the key**. Not by whoever holds the reference.
+
+There is no third party in that sentence. No server, no ACL, no check at access time: the engine verifies a permission on WRITE only (`verify_permission` → `PermissionDenied` in `Commit::verify`), never on read. Content is encrypted; the key IS the right. That is what local-first means here — there is nobody to ask, so possession is the whole mechanism.
+
+### Why this is easy to lose
+
+Client-server reflexes supply the missing half without asking. In that world a link is an identifier, harmless on its own, and access is decided elsewhere — so handing out a URL feels safe, and "may Bob read this?" is a question with an answer. Both instincts are wrong here, and neither announces itself: code written on them **does not fail**, it silently grants. There is no error to notice, no test that goes red on its own.
+
+### What the rule buys — composable confidentiality
+
+Because a reference grants nothing, **a reference is not recursive**: a widely-circulated document may point at a restricted one. The reference says *it exists*; the key says *who reads it*.
+
+That is what lets confidentiality be composed inside one document:
+
+- a public summary pointing at restricted chapters,
+- a public event pointing at its participant list,
+- a profile pointing at data reserved for a circle.
+
+The author circulates one document and still decides, per referenced document, who gets the key. **Remove non-recursiveness and this whole construction collapses**: everything a circulated document mentions becomes readable by all its readers, and no level of confidentiality can exist inside a document one shares.
+
+`test/cross-user-access.test.ts` is this property's test, and it is the suite's centrepiece: Alice's document in a public store REFERENCES her protected one; Bob follows the reference, names the target, reads nothing of it. It writes the reference **bare** — which is exactly what an application must do.
+
+### The failure mode to watch for
+
+Any call that returns a key where a reference was asked for. It looks like a convenience ("give me the link to my document"), it type-checks, the tests stay green — and it converts "whoever has the reference and the key" into "whoever has the reference", for that document and every document it mentions. `linkTo` (added and removed 2026-08-06) was exactly this.
+
+---
+
 ## 1. A ReadCap = possession of a key, NOT a per-identity ACL
 
 A **ReadCap is fundamentally a cryptographic key that one holds**, not an ACL entry tied to a wallet. "Whoever holds the key can read."