chore(ci): tune PR Agent config

- split per-model banners so two model jobs no longer overwrite each other
- raise reviewer findings cap to 8, broaden /improve to readability/cleanup
- enable dual-publishing for high-score suggestions (inline annotations)
- switch Claude model from opus-4-7 to opus-4-6 (fallback sonnet-4-6)
- raise reasoning_effort to high, response_language to zh-CN
- drop two dead config keys silently ignored by upstream schema
- add best_practices.md with 6 project-specific invariants for /improve

.pr_agent.toml

@@ -1,22 +1,41 @@
+# PR-Agent configuration for InfiPlot.
+# Per-model banners and tool-specific extra_instructions live in
+# .github/workflows/pr-agent.yml. Project-level coding rules live in
+# best_practices.md (consumed by /improve). Keep this file structural only.
+
 [config]
 ai_timeout = 300
 temperature = 0.2
+reasoning_effort = "high"
+response_language = "zh-CN"
 
 [pr_reviewer]
-num_code_suggestions = 4
-inline_code_comments = true
+num_max_findings = 8
 require_security_review = true
-extra_instructions = """
-This is a Next.js 16 / React 19 / TypeScript interactive visual novel engine.
-Focus on: logic errors, security vulnerabilities, missing error handling,
-type safety issues, and architectural violations.
-Do not comment on code style or formatting.
-"""
+require_tests_review = false        # repo has no test framework
+require_can_be_split_review = true
+require_score_review = true
+require_todo_scan = true
+persistent_comment = false          # two model jobs would otherwise overwrite each other
 
 [pr_description]
 generate_ai_title = true
 publish_labels = true
 use_bullet_points = true
+enable_pr_diagram = true
+
+[pr_code_suggestions]
+num_code_suggestions_per_chunk = 6
+max_number_of_calls = 4
+parallel_calls = true
+focus_only_on_problems = false       # also surface readability/reuse cleanups, not only bugs
+suggestions_score_threshold = 0
+dual_publishing_score_threshold = 7  # high-scoring suggestions also appear as inline comments
+persistent_comment = false           # two model jobs would otherwise overwrite each other
+
+[best_practices]
+enable_global_best_practices = false
+max_lines_allowed = 2000
 
 [ignore]
 glob = [