fix: recover text tool-calls and keep transient noise out of history

[jbj338033]

Summary

Transient model/runtime anomalies — a tool call emitted as plain text, a machine status line, an empty reasoning block — were persisted verbatim into the conversation and re-sent, so a one-off error became a self-reinforcing, cache-amplified failure that even a direct human correction could not overcome. This makes the harness robust to those anomalies:

• Recover tool-call-as-text: the engine detects a tool call the model emitted as assistant text, recovers it into a real tool call (dedup'd against any structured call), executes it, and strips the leak from persisted/resent history so it cannot snowball. Provider-agnostic policy in the engine; dialects (<invoke>/antml: universal, Hermes/Mistral for local models) in a new tool_recovery module.
• Round-trip reasoning: preserve thinking blocks that carry only a signature (empty summarized text) so the model's reasoning anchor survives across rounds.
• Provenance: label background-process output/roster as environment notices rather than synthetic user messages, so the model stops replying to and repeating them.

Verification

• cargo fmt --all --check — passed
• cargo clippy --workspace --all-targets -- -D warnings — passed
• cargo test --workspace — passed (goat-agent 125, incl. 13 tool_recovery unit tests over the real leaked sample and local dialects)