Add safe tool audit events

[chubes4]

Summary

• Add tool_audit_events to mediated conversation loop results as a safe replay/audit surface that omits raw parameters while preserving deterministic hashes.
• Classify core mediated tool failure paths (tool_not_found, missing_required_parameters, executor_exception) and validate the audit event contract.
• Document the audit event shape and redaction filter for host-specific deterministic parameter normalization.

Context

Dan Luu's wp-gym review exposed that downstream benchmark consumers need replayable tool-action evidence without each corpus inventing its own episode-step shape: Automattic/wp-gym#76

This addresses the Agents API foundation issue: Closes #170

Testing

• php tests/conversation-loop-tool-execution-smoke.php
• php tests/conversation-loop-smoke.php
• php tests/conversation-loop-transcript-persister-smoke.php
• php tests/tool-runtime-smoke.php
• for f in tests/*-smoke.php; do php "$f" || exit 1; done
• php -l src/Runtime/class-wp-agent-conversation-loop.php
• php -l src/Runtime/class-wp-agent-conversation-result.php
• php -l src/Tools/class-wp-agent-tool-execution-core.php

AI assistance

• AI assistance: Yes
• Tool(s): OpenCode (GPT-5.5)
• Used for: Drafted the implementation, tests, and documentation; Chris directed the upstream factoring and remains responsible for review and merge decisions.