[ethan-scitix]

Type

• chore — CI, tooling, dependencies, config

Summary

• Release 0.15.1: CHANGELOG entry and Dockerfile wheel version, as in the 0.15.0 release (chore: release 0.15.0 #260).
• A patch release, nothing breaking: the vllm_native dialect (feat(models): add the vllm_native dialect (RFC #25's seventh) #261), the google_interactions dialect (feat(models): add native Google Interactions support #263), a scaled SGLang group that launches with SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN=1 (fix(cli): let a scaled SGLang engine serve past its derived window #265, closes Scaled SGLang window fails at startup when the YaRN block names original_max_position_embeddings #264), and a corrected 0.15.0 migration note (docs(changelog): name no private loader as the 0.15.0 replacement #262). No "Migrating" section.
• On disk: IDENTITY_SCHEMA stays 5 and RESULT_SCHEMA_VERSION stays 1 (neither line changes in v0.15.0..main). feat(models): add native Google Interactions support #263 makes request_definition read an HTTP envelope's body.stream; no existing digest changes, because GenerateContent refuses stream in its body (an IR-owned key) and still derives streaming from path and query. fix(cli): let a scaled SGLang engine serve past its derived window #265 sets the variable on the launch command only, so neither infer_plans*.yaml nor the inference identity changes.
• Trade-off carried from fix(cli): let a scaled SGLang engine serve past its derived window #265: on a scaled SGLang group, context.max is now the only bound on the window.

Test Plan

Automated

• ruff check ., ty check: clean
• pytest tests/unit tests/integration tests/acceptance -m "not benchmark" -n auto: 10897 passed, 94 skipped, 1 failed. The failure is test_exec_quotebench_anchor.py on a macOS host (BSD mv has no -T; it compares against GNU verdicts), as on 0.15.0.
• Benchmark gate pytest -m benchmark -s, serial, no coverage: all scenarios PASS (single-turn 483.7 SPS / 75.6%, long output 161.5 / 75.7%, multi-iteration 171.0 / 80.2%, high concurrency 849.3 / 66.4%, framework overhead 739.0 SPS, resume 1.0 s, peak RSS 422.6 MB)
• pre-commit run -a: all hooks pass
• check_preflight.py --level deep: 43 pass; one fail, check_version (latest tag 0.15.0 vs CHANGELOG 0.15.1, resolved by tagging v0.15.1 after merge); link warnings for placeholder hosts, provider roots, not-yet-existing compare link and two connection timeouts.

Checklist

Required (all PRs)

• PR title follows conventional format
• No internal paths, credentials, or personal info in committed files

🤖 Generated with Claude Code