[Leon-Algo]

Summary

• Versioned prompt profiles: graphgen/templates/prompt_profiles.py + BaseGenerator.TEMPLATE_KEY/apply_prompt_profile/template(); all 11 generator build_prompt overrides become instance methods. Profiles are JSON directories structurally validated against the official template constants (fail-closed); built-in example.v1 ships byte-identical official templates as an example. No profile applied → byte-identical official behavior.
• Generic output-gate registry: graphgen/operators/generate/gates.py with register_gate/GATES and a generic non_empty_answer gate. No domain-specific gates included.
• Bounded generation concurrency: event-loop semaphore exposed as GenerateService(llm_concurrency=N); default unchanged.
• GenerateService opt-in options: prompt_profile / output_gate / llm_concurrency — all default to upstream behavior.
• Env-gated LLM client patches (all off by default): GRAPHGEN_EMPTY_RETRY bounded empty-response retry (rate-limiter re-applied per attempt), GRAPHGEN_EMPTY_EXTRACTION_PREFILL JSON-compatible assistant prefill { for the schema-guided extractor (the previous tuple-format prefill can never parse against json.loads), GRAPHGEN_DISABLE_THINKING for logprob calls, and GRAPHGEN_LLM_ACTOR_MAX_CONCURRENCY Ray actor option + logger-context fallback.
• Packaging: package_data gains templates/profiles/*/*.json only. Upstream package identity (name/version/url/author) untouched.

Second of three PRs split from #192 per review; no PMS assets, no fork identity. PR A (partitioner test fixes) and PR C (event-entity pipeline, builds on this) follow separately.

Validation

• pytest tests/unit_tests → 20 passed (profiles 8, gates 3, llm_concurrency 2, openai patches 4, log 1, plus collection checks).
• pylint --rcfile=.pylintrc baselines/ graphgen/ webui/ → 10.00/10, zero messages (CI gate green; resolves the arguments-differ findings on all generator overrides noted in Add versioned prompt profiles and event-entity generation pipeline #192).

🤖 Generated with Claude Code

[ChenZiHong-Gavin]

LGTM, merging. Checked that the default path stays byte-identical (template() returns the same official constant objects; all build_prompt callers already use self.), the RPM/TPM wait now sits inside the retry loop, and pylint is green on 3.10–3.12.

Two non-blocking nits for a follow-up:

1. MaskedFillInBlankGenerator.TEMPLATE_KEY = "masked_fill_in_blank" but profiles/example.v1/ has no masked_fill_in_blank.json, so applying the shipped example profile to that generator raises prompt_profile_template_missing. Either add the file (same content as aggregated.json) or point the key at aggregated.
2. In SchemaGuidedExtractor, the prefill retry passes assistant_prefill="{" but the response is fed to json.loads as-is. A backend that genuinely continues the assistant turn returns the text after {, so the retry still cannot parse; prepending the prefill when the response does not already start with it would make this robust for both backend behaviours.

[Leon-Algo]

Thanks for the quick review and merge, @ChenZiHong-Gavin! We really appreciate you double-checking that the default path stays byte-identical — that kind of care for backward compatibility is one of the things that makes GraphGen such a solid project to build on. Thanks also for the two nits you spotted in this review; they pointed at real gaps, and we have follow-up fixes for both (see #197).