fix(shepherd): correct the settle procedure and stop restating its command

AGENTS.md

@@ -1,3 +1,3 @@
 # AGENTS.md
 
 Guidance for AI coding agents (Claude Code, Gemini CLI, GitHub Copilot, Codex,
@@ -324,17 +324,23 @@
   A badged finding stated **outside an inline thread** — in a top-level
   comment or in a review's own body — has no reply linkage, so the
   reply-based adjudication path cannot reach it and findings outrank a later
-  clean result on the same head. Record its disposition with the checker's
-  `settle` subcommand: answer the finding on the PR as usual (fix it, decline
-  it with evidence, or file it), then `settle --surface comment|review --id N
-  --disposition declined|filed --note …`, adding `--covers <n>` where the
-  target states more than one finding. `check` then treats that finding as
-  answered and the cycle can reach terminal-clean, reported with a detail
-  naming the disposition. The record is durable and head-bound: it is
-  fingerprinted against the body it settled, so a finding Codex edits
-  afterwards blocks again, and any push starts a fresh cycle as usual.
-  Settling is not a substitute for answering — it records an adjudication a
-  human wrote, and the note is required for that reason.
+  clean result on the same head. The checker's `settle` subcommand records the
+  disposition instead; its exact invocation lives with the recipe in
+  `ai/skills/universal/shepherd/SKILL.md`, which this file deliberately does
+  not restate — the same reason it routes you to the checker rather than
+  describing how to poll.
+  What belongs here is when it applies. Answer the finding on the PR as usual,
+  and note that only two of the three answers end with `settle`: **fixing** it
+  means a push, which moves the head and starts a fresh cycle that reviews the
+  fix on its own merits, and `settle` neither applies nor accepts that
+  disposition. It records the two answers that leave the code alone —
+  declining with evidence, or filing it as follow-up work — and requires a
+  note for exactly that reason: the record is a human's adjudication, not a
+  suppression. It is head-bound and fingerprinted against the body it settled,
+  so a finding Codex edits afterwards blocks again while the superseded entry
+  survives as the record of what was decided about the earlier text. Once
+  every non-thread finding on the head carries one, `check` reports clean with
+  a detail naming the disposition applied.
   Shepherd is **externally driven** — CI results and other people's comments
   are its input, so it cannot manufacture a round on its own. A round is one
   fix push, or one no-change cycle where everything is answered and nothing

ai/skills/universal/shepherd/SKILL.md

@@ -700,8 +700,12 @@ you rather than the bot):
   reply to, so no act on GitHub can ever record that you answered it and
   `check` returns exit 10 for that head forever — the deadlock the inline
   adjudication path was built to end, reappearing on the two surfaces it
-  cannot see. Answer the finding on the PR as usual (fix it, decline it with
-  reasoning, or file it), then record the disposition:
+  cannot see. Answer the finding on the PR as usual — and note that only two
+  of the three answers end here. **Fixing** it means a push, which moves the
+  head and starts a fresh cycle that reviews the fix on its own merits;
+  `settle` neither applies nor accepts that disposition. For the two answers
+  that leave the code alone — declining with reasoning, or filing it as
+  follow-up work — record the disposition:
 
   ```bash
   "$helper" settle --state "$state" --actor-id 199175422 \