Release v0.6.0: expanded real-model evidence, adapter lint, reporting, governance

[sergeyizmailov]

Summary

Release v0.6.0 — consolidated release PR (parallel reviewer + this pass, all verified locally).

Real-model evidence (Priority 0/3 of the pre-submission review):

• Reviewed Llama 3.3 70B expanded-suite report: 80 trials across the four representative pairs — 30/40 attacks produced the expected violations (V1/V2 10/10, V6-as-V2 10/10, V7 10/10; V3 not exercised on the first prompt), 0/40 benign false positives, 40/40 benign task success, 0 errors.
• Separately preserved 20-run V3 follow-up with the strengthened prompt: 10/10 expected V3 violations, 10/10 benign clean.
• Qwen expanded-suite first attempt is not published (4 attack-trial timeouts + harness_commit misattribution in the pre-fix harness); it is preserved privately and the rerun is one dispatch of the now per-scenario real-model-benchmarks workflow. Every published report carries a review block with source hash and an explicit corrections log; per-run records are never modified.

Feature and hardening changes (already reviewed in #31/#32/#34): validate-adapter linter, --junit-detail, contract-validation wording, GOVERNANCE.md, ruff baseline pinned + CI gate, dependabot cooldown, harness hardening (HARNESS_COMMIT at process start, endpoint URL validation, progress flushing), benchmark review protocol, per-scenario benchmark workflow with concurrency control.

Version: 0.6.0 across pyproject.toml, __init__.py, CITATION.cff, CHANGELOG (0.6.0 section).

Verification

• 364 tests pass, corpus 75/75 in all defense modes, ruff check . clean, actionlint clean, repo-wide zizmor clean, git diff --check clean, wheel builds as delegationbench-0.6.0.
• Suite aggregates independently recomputed from raw run records before commit.
• No credentials in any artifact; endpoint redacted.