docs(adr): ADR 0021's D57–D63 as built

How the budgets, handles, the operator's screens and effort read where the
decisions left something open: optional limits, money stopping only
priced assistants, a handle typed whole as one suggested, 50 at a time
with no search where lists stay short, lists changed in place after an
action, the effort's ceiling on output, and own keys without effort.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>

docs/adr/0021-private-instance-and-assistant.md

@@ -405,6 +405,41 @@ assistants' effort work:
   one above the highest. Effort spends more tokens at the same price per token, so the
   budget (D58) still bounds what it costs.
 
+As built on 26 Oct, for v0.22.0 (build 25, which the server needs: `minimumAppBuild` 25,
+as older apps read limits per grant), D57–D63 read so where they left something open:
+
+- **Budgets** (D58, D59): each of the four limits may be empty, and all four empty
+  leaves the person no budget; granting someone who has one replaces it, which the sheet
+  says first, naming whose. Money stops only an assistant with prices: one without counts
+  its tokens, which the token limits stop, and the sheet offers money limits only while an
+  assistant set up has prices. A change that leaves out an assistant no longer set up
+  drops its grant, which the sheet says.
+- **Handles** (D61): Grant and Appoint take a handle picked among those suggested, or
+  typed whole as an account's (Enter picks it); any other is refused before anything is
+  asked.
+- **The operator's screens** (D62): read 50 at a time from where the list stands, the
+  next 50 once the person scrolls to within 600 pixels of its end. Accounts and the
+  assistants' people are searched by their handle's start; Waiting and Operators aren't,
+  as at most 10 accounts wait at once (point 3) and operators are few. After an action the
+  list changes in place (approved, refused and taken back leave it; suspended and named
+  change their row; a changed grant reads its person anew) rather than reading everything
+  again.
+- **Effort** (D63): `_MAX_EFFORT` without `_EFFORT` stops the server as it starts, as any
+  other setting it can't read. A question asks for its effort's ceiling on output, which
+  the instance holds it to (`low` 8,192 tokens to `max` 65,536; the model's own effort
+  16,384), since the thinking counts towards it. The level chosen in the field is the
+  person's for that assistant on that device (`askEfforts`): a conversation continued from
+  the history asks at it too, while the flashcards proposed, Explain and Why? ask at the
+  assistant's own, as nobody chose one there. A person's own key gets no effort, as its
+  provider's levels aren't known, and keeps the ceiling of 4,096 tokens it had.
+- **The migration** moves each person's grants' limits into one budget (per unit, their
+  sum; a grant without a limit leaves that limit empty) and then drops the grants' three
+  limit columns. Unlike Learn's other migrations, it doesn't keep the old server working
+  until the switch: in the minute between them, the old server's assistant calls fail
+  (asking, granting, the figures), everything else serving on. The developer accepted
+  this (26 Oct), as assistants went live that day with one person using them, over
+  keeping the columns unused for a release and dropping them in the next.
+
 ## Proposal
 
 1. **Who signs up, per instance:** `SIGN_UP`: `open` (as today), `approval`, or