docs: add pull request template

.github/pull_request_template.md

@@ -0,0 +1,56 @@
+<!--
+Write for a reviewer, not as a work log.
+
+Completion criterion — post only when every item is true:
+
+1. The opening is one sentence naming the concrete change and what it is.
+2. A runnable consumer example follows immediately. It defines every input,
+   uses real imports and values, and explains domain terms in plain language.
+3. Every changed public entry point is demonstrated through usage. Add more
+   code, output, or tables when one example cannot cover the whole surface.
+4. Risk appears near the top and states what could break and what needs
+   reviewer judgment.
+5. Every narrative heading states a fact specific to this change.
+6. Code, output, or a table appears before prose that explains it.
+7. Verification, findings, and declined alternatives live in <details>.
+8. Coverage changes state the numeric delta and its cause, including whether
+   scenarios were added, removed, or reclassified.
+9. Every sentence passes the reuse test: if it could appear unchanged in a
+   different PR, cut it.
+10. Replace every placeholder and delete this instruction comment.
+
+State mechanisms precisely with the codebase vocabulary. Keep each meaning in
+one place.
+-->
+
+REPLACE: One sentence naming what this PR changes and what it is.
+
+```ts
+// REPLACE: Runnable consumer code that defines every value it uses and shows
+// the result. Change the fence language when TypeScript is not appropriate.
+```
+
+## Risk
+
+<!--
+REPLACE: State what could break, what needs judgment, and why the risk is
+bounded. Delete this section only when there is genuinely no risk.
+-->
+
+## REPLACE: A heading that states a fact about this change
+
+<!--
+REPLACE: Show code, output, or a table first. Follow it with only the explanation
+the example cannot carry. Repeat fact-based sections until every changed public
+entry point has been used.
+-->
+
+<details>
+<summary>Implementation notes and verification</summary>
+
+<!--
+REPLACE: Record findings, decisions declined, commands, exit codes, test counts,
+and coverage deltas with their causes.
+-->
+
+</details>