fix: une barriere appartient a la tentative qui l'a produite

Seconde revue adverse, sur le correctif precedent - et elle a raison :
celui-ci MASQUAIT la barriere perimee au lieu de l'oublier.

`reportSetupFailure` effacait `establishing` et laissait `lastState`. Le
premier correctif avait ajoute « et un canal tient » du cote du LECTEUR,
ce qui ne cachait la barriere morte que tant que rien ne tenait. La
souscription suivante sur le meme fan-out remettait `establishing` a
vrai, le masque tombait - et le fan-out rejouait ce `State` au nouvel
arrivant. Mesure : l'ouverture bootstrap franchissait sa barriere en 2 ms
sur un `State` d'un appel qui avait ECHOUE, journalisait
`BARRIER ... synced`, et `getSyncState` repondait `"synced"` - donc
autorite, donc `noRowsMeansAbsent` prenait une lecture a zero ligne pour
une absence definitive. Le depot detruit que le commit precedent
pretendait empecher.

Corrige a la source : une tentative qui echoue oublie sa barriere, comme
le font deja `releaseFanOut` et `resubscribeDocs`. Le garde du lecteur
est RETIRE avec, et pas garde en ceinture-bretelles : il etait devenu
inatteignable, et c'est lui qui avait cache la faute. Apres : la barriere
est oubliee, l'ouverture attend la sienne (202 ms), `"timed-out"`.

Et le test que j'avais ecrit affirmait le defaut - `barrierCrossed`
vrai - au lieu de l'etat correct. Corrige, plus un second test qui
conduit le rejeu de bout en bout.

packages/polyfill/src/surface/subscribe.ts

@@ -299,6 +299,14 @@ export function docChannel(nuri: Nuri): DocChannel {
   // two together are the only thing that separates "asked" from "open" - and recording the
   // first as if it were the second is the defect this accessor exists to make impossible.
   const state: DocChannelState = !entry.establishing ? "none" : entry.realUnsub !== null ? "open" : "opening";
+  // `lastState` alone, and deliberately NOT `&& state !== "none"`. Every path that stops an
+  // attempt standing forgets the barrier with it - {@link releaseFanOut}, {@link resubscribeDocs}
+  // and, since this was found the hard way, {@link reportSetupFailure} - so the two can no longer
+  // disagree, and a guard here would be unreachable. It would also be a MASK: the first repair of
+  // that defect put exactly this condition in the reader instead of forgetting the barrier at the
+  // source, and it hid the fault while the fan-out went on replaying a dead `State` to the next
+  // joiner. A fact that is wrong at the source has to be fixed at the source; a reader that
+  // papers over it only moves the lie.
   return { state, barrierCrossed: entry.lastState !== null, incarnation: entry.incarnation };
 }
 
@@ -383,6 +391,15 @@ function reportSetupFailure(nuri: Nuri, entry: DocFanOut, error: unknown, attemp
   console.error("[subscribe] doc_subscribe failed for", nuri, error);
   if (!isCurrentAttempt(nuri, entry, attempt)) return;
   entry.establishing = false;
+  // A BARRIER BELONGS TO THE ATTEMPT THAT PRODUCED IT, and this one failed. The push may well
+  // have arrived - the callback is driven locally while the call's own answer travels the RPC -
+  // but the call did not take, so there is no subscription behind that `State` and nothing will
+  // follow it. Left here, the fan-out replayed it to the next joiner: a bootstrap open crossed
+  // its barrier in 2ms on a `State` from a call that had FAILED, logged `BARRIER ... synced`,
+  // and `getSyncState` answered `"synced"` - which is authority, so a zero-row read over the
+  // document became a definitive absence and a deposit was destroyed. {@link releaseFanOut}
+  // and {@link resubscribeDocs} both forget it for the same reason; this path did not.
+  entry.lastState = null;
   for (const listener of [...entry.listeners]) {
     const tell = listener.onSetupFailed;
     if (tell === null || !entry.listeners.has(listener)) continue;