fix(code-review): stop the inline pass reading as the session's final output

plugins/tend-ci-runner/skills/code-review/SKILL.md

@@ -94,3 +94,5 @@ On a core-logic change, take one more pass as a fresh reviewer holding the verif
 Return the findings to the caller as a list of at most 10 (at most 4 for a peripheral change), ranked most-severe first, each with `file`, `line`, `summary`, `failure_scenario`, and its verdict. If nothing survives verification, say so in one line. Don't call the `ReportFindings` tool and don't publish an artifact — the caller owns the output.
 
 Tell the caller which mode ran (fanned-out angles with a subagent verify, or a single inline pass) so it can weigh the findings — context for the caller, not content for the review it posts.
+
+**Running inline, this is a sub-step, not the answer.** When the pass runs in the caller's own session rather than a subagent, there is no separate caller to return to — "return the findings" means hand them to the caller's next step and continue there. A session that emits them as its final message ends with the caller's work unfinished and nothing posted, leaving the findings reachable only from the session log.

plugins/tend-ci-runner/skills/review/SKILL.md

@@ -193,7 +193,7 @@ Scale its depth to how core the change is:
 - Peripheral or mechanical (config, dependency bumps, test-only, docs that don't assert how the code behaves): tell it the change is peripheral, so it runs the short angle set in one pass.
 - The project's core logic, or prose asserting how it behaves: tell it the change is core, so it fans the angles out and sweeps for gaps. Prose is checked by reading the code it describes, so a one-line Markdown diff can still be core.
 
-What counts as core is repo-specific; let the project's own guidance (CLAUDE.md, a repo review skill) or your judgment decide. Both passes feed one verdict: fold its findings into the review you submit in step 6. It only reports back — it never posts a review, comment, or commit of its own, so the dedup and single-review path is preserved.
+What counts as core is repo-specific; let the project's own guidance (CLAUDE.md, a repo review skill) or your judgment decide. Both passes feed one verdict: fold its findings into the review you submit in step 6. It only reports back — it never posts a review, comment, or commit of its own, so the dedup and single-review path is preserved. Its findings are not the review: when it returns, continue to step 6.
 
 ### 6. Submit