fix(conversation): cover validation retries and stale mute state

[jakepu]

Summary

This is a focused follow-up to NousResearch/hermes-agent#63888. It keeps that PR's original fix and regression test, while closing three adjacent state-machine gaps found during review.

Changes

• Move housekeeping/substantive classification immediately after tool-name repair, before validation branches can retry or continue.
• Treat unknown and disabled tool names conservatively as substantive.
• Clear _mute_post_response together with the cached fallback fields when substantive work resumes.
• Hoist the post-response housekeeping policy to a documented module-level constant.
• Add parameterized regression coverage for:
  • valid substantive tool-only turns
  • malformed, non-truncated JSON arguments
  • unknown tool calls
  • mixed housekeeping and substantive calls
• Add positive coverage for legitimate same-turn housekeeping fallback behavior.
• Add stream-consumer coverage proving substantive work unmutes progress output.

Why

On the current NousResearch#63888 head, the expanded suite produced:

6 focused scenarios: 3 passed, 3 failed

Malformed JSON and unknown-tool retries could bypass classification and still finalize two-turn-old housekeeping narration. A normal substantive turn also left _mute_post_response=True, suppressing progress output.

With this patch, all seven focused tests—including NousResearch#63888's original standalone regression—pass.

Verification

Focused fallback tests:                  7 passed
Full run_agent test file + original test: 423 passed
Related finalizer/state tests:            6 passed
Ruff:                                     passed
Python compilation:                       passed
git diff --check:                         clean

This PR targets the existing contributor branch so the improvements can be incorporated into NousResearch#63888 without opening a competing upstream PR.