fix(ai): read a Codex sign-in's memory only under its own account

Every write to the per-sign-in Codex memory checked the account digest,
but the reads did not. A credential reconnected to another ChatGPT
account therefore tried the old account's working model first, sorted
by its refusals and trusted its model list, until the first write
rebound the entry, and the vision gate kept answering from the old
account in the meantime.

The client reads (working model, refusals, model list) now take the
credential and treat an entry learned under another account as empty.
The vision gate holds no account id, so the connect routes rebind the
entry and the disconnect routes drop it, and the gate reads the entry
as its process last bound it.

docs/codex-protocol-spec.md

@@ -692,7 +692,7 @@ The chain is overridable via the env var `CODEX_MODEL_FALLBACK_CHAIN` (comma-sep
 | `401` after refresh, `403`, `429`, `5xx`                                      | propagate (don't walk)                           |
 | `200` with SSE `response.failed.error.code === "invalid_prompt"`              | propagate (request shape is wrong, slug is fine) |
 
-**Positive cache** — when a slug is observed working (HTTP 200 with at least one `response.completed` event), record `(slug, timestamp)` for that sign-in. On its next request, jump straight to the cached slug; the model list is read only when there is none or it stops working. TTL: 1 hour. Scope: per stored credential (`user:<id>` for a person's connection, `central` for the operator-shared one), process-local, no DB write, at most 1000 sign-ins held. Each entry carries a digest of the ChatGPT account id it was learned under, and a credential reconnected to another account starts empty. Two sign-ins on different plans no longer overwrite each other's working slug.
+**Positive cache** — when a slug is observed working (HTTP 200 with at least one `response.completed` event), record `(slug, timestamp)` for that sign-in. On its next request, jump straight to the cached slug; the model list is read only when there is none or it stops working. TTL: 1 hour. Scope: per stored credential (`user:<id>` for a person's connection, `central` for the operator-shared one), process-local, no DB write, at most 1000 sign-ins held. Each entry carries a digest of the ChatGPT account id it was learned under, and a client reads it only under that account, so a credential reconnected to another account starts empty: its first request tries no slug, skips no refusal and trusts no list the old account left, in any process. The vision gate holds no account id; the connect and disconnect routes rebind or drop the entry, so in the process that handled the change it answers from the new sign-in at once, and elsewhere from that process's first request on it. Two sign-ins on different plans no longer overwrite each other's working slug.
 
 **Cache invalidation**: on a 400/404 walk the refused slug leaves the working slot and goes to the back of that sign-in's line for six hours. The next-best slug is promoted into the cache only after it succeeds.