chore(release): start 0.6.6 development

[bingran-you]

Summary

• move main from the published 0.6.5 final version to 0.6.6.dev0
• refresh uv.lock
• restore the internal-preview version line after the public release

Validation

• tools/release_version.py internal-preview --run-number 123 -> 0.6.6.dev123
• pytest tests/test_release_version.py -> 22 passed

[bingran-you]

Daily scan triage (2026-07-11): release-bump PR is clean and GitHub checks are green (test, pip-audit, manifest-parity, integration-light / rollout-smoke; integration-scope jobs skipped/cancelled as expected for the light-scope diff).

Holding unmerged under repo policy: this still needs non-author human review before main because self/bot review is not enough. Added review:pending; leaving status:ready.

[bingran-you]

Users Simulation automation review (2026-07-11): ready by simulation on head df51c2d660b055b3d399a5dc0e3ebb87c922d21d.

Scope checked: release-start 0.6.6.dev0 versioning, package metadata, CLI/SDK surface, release helper behavior, and deterministic artifact-shape tests.

Evidence:

• Diff is only pyproject.toml + uv.lock: 0.6.5 -> 0.6.6.dev0.
• GitHub current-head checks are green: test, pip-audit, manifest-parity / parity, and integration-light / rollout-smoke passed.
• Local smoke passed: uv sync --extra dev --locked, ruff check pyproject.toml, ty check src, bench --help, and SDK Scene.single(agent="oracle").
• Subagent package/release checks also passed: uv lock --check, release-version tests, CLI misc/dataset registry tests, wheel/sdist build + metadata check, clean-wheel install, and artifact-shape tests.
• SDK version smoke reported benchflow.__version__ == 0.6.6.dev0; tools/release_version.py internal-preview --run-number 424242 produced 0.6.6.dev424242.

Artifact health:

• Current-head integration-light run 29141635529 validated healthy with benchflow-experiment-review: 1/1 healthy, ACP + LLM trajectories present, training-ready results.jsonl, reward present, 320,708 tokens, 17 tool calls, 255.8s timing.
• Caveat: this GLM user-endpoint artifact still has null agent_result.cost_usd / price_source; token, timing, and trajectory metadata are complete.

Thermo review: no structural blocker. This PR adds no runtime branch, no wrong-layer logic, no special-case condition, no abstraction, and no file-size regression attributable to the PR.