0.1.0

First version, built and tested against DeepSeek Harness 0.2.1-rc.2, and also checked on 0.2.1-rc.3.

• Audits a finished turn at agent/turn-stopping: rebuilds the turn's tool calls, tool output, and final text from the
  Session events, gates the turn, and asks a model auditor (thinking off, a facts-and-limits card, a hard output cap).
• Routes by severity: a contradiction that cites text found in the evidence steers a short correction, so the model
  revises in the same turn; an overstated or unsupported claim becomes a flag.
• Free scanner in front: absolute words, and numeric claims (numbers with units, version strings, ISO dates) that the
  tool output does not contain. Turns that ran no tool never reach the model; their absolute words become a scanner flag.
• One audit per turn (the hook fires again after a steered correction).
• Flags go to a plugin:second-angle/verdict Session record, one stderr line, and the harness logger, and optionally
  to a notice that reaches the model and the chat with the next turn.
• sync and async modes, a latency bound that holds even when a provider ignores the abort, and fail-open
  behavior on every auditor error, timeout, or empty output.
• A public eval set of 24 cases built from the harness's own documentation, with a runner and a scorer.

sha256 of dsh-second-angle-0.1.0.tgz: 3217b294b5ab01c142d11b8b817d58f0457fbe445f57f073b1a10e969bb65b93