refactor(naming): use semantic publication and batch identifiers

[seonghobae]

Scope

This is the canonical naming repair for two private contextual-orchestrator boundaries: NIM benchmark-evidence atomic publication and the synchronous adapters over injected async pg-llm-batch clients. The branch has been reconciled against the current main@464da4715b495b5eaaa593eba3796e2d976ee0c9; external/public contracts remain unchanged.

TDD lineage

NIM evidence publication

• RED 6ea1f728e267ce9983a1a599c34c8989cae0a52e requires _publication_residue_paths(final_directory, residue_kind) and proves the old _residue authority must disappear.
• Repair b2ccbd663e6c624d1b0773bf41409ced0f7f840e applies the semantic rename through all private callers and co-located publication-state identifiers.

Old → new vocabulary includes _residue → _publication_residue_paths, final → final_directory, kind → residue_kind, prefix → residue_prefix, flags → open_flags, descriptor → lock_descriptor, staging → staging_directory, backup/backups → backup_directory/backup_paths, and artifact name/payload → artifact_name/artifact_payload.

pg-llm-batch sync/async adapter

• RED 21d1c18d91497730cfc887afa353084359f92322 adds an AST/source contract for the two pg-llm-batch adapter classes. The then-current production file still exposed two private _run(coro) methods and six self._run(...) callers, so the regression corresponds to the exact current defect.
• Repair 36996df189bd40e3d7223a710d83edc7c70c8eba renames both private bridges to _run_async_client_operation(async_operation) and propagates all six submit/poll/retrieve callers.
• Cleanup 0fbb183a3151a2674a40aa8ae0d58eb74defbc97 restores two pre-existing docstring literals that were accidentally escaped while replacing the large source file. Comparing the RED head to this exact head leaves only the intended 10 additions/10 deletions in batch_routing.py.

Compatibility and persistence

No public Python call shape, OpenAI-compatible JSONL field, pg-llm-batch client method, serialized NIM evidence schema, artifact filename, provenance field, database object, migration, FK, index, ORM mapping, UPSERT path, partitioning, lock model, or read/write topology changes. id, status, body, data, and similar fields required by upstream/OpenAI-compatible payloads remain at their protocol/adapter boundary rather than being mechanically renamed.

Verification boundary

Current exact head: 0fbb183a3151a2674a40aa8ae0d58eb74defbc97. Keep Draft until fresh exact-head hosted checks are terminal-success, current review findings/threads are resolved, and a qualifying independent non-author approval applies to the unchanged last push. Predecessor/base/model-only evidence is not merge evidence; do not self-approve, force-push, or bypass protection.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}