pawbot v0.7.2

pawbot v0.7.2

This release adds DeepSeek-focused model and output handling, safer recovery
from truncated tool calls, and WebUI controls for turning saved regression
samples into task evaluation cases.

Changes

• DeepSeek Flash and DeepSeek V4 Pro are included in the capability catalogue;
  legacy DeepSeek V4 Flash identifiers remain compatible aliases.
• New DeepSeek model configurations default to a 65,536-token output budget;
  other providers keep their existing defaults.
• Truncated tool-call arguments are rejected before execution. The Agent gets
  one strategy-changing recovery attempt, then receives a structured error if
  the same output limit is hit again.
• Repeated safe-to-retry tool failures are stopped with a structured
  no-progress error instead of consuming the rest of the turn budget.
• WebUI candidate runs display conversation summaries, and saved regression
  samples can be added to or removed from the Agent task evaluation set.

Provider compatibility

DeepSeek is the most thoroughly validated provider path in this repository.
Other providers continue to use the shared adapter. Provider-specific tuning,
tests, issues, and pull requests are welcome, especially with sanitized
reproductions.

Verification

The release candidate passed the complete GitHub Actions matrix and local
Python, WebUI, type-checking, build, package, and provider-free evaluation
checks.