fix: use clean user message for auto-title instead of skill-bloated input

[dieutx]

Summary

Sessions started via skill invocation get meaningless auto-generated titles like "Skill Invocation and Content Loading" instead of the user's actual intent. Follow-up to #4940 which fixed the same issue for memory providers.

Root Cause

When a skill is active, the `message` variable contains the full skill activation payload: a `[SYSTEM: The user has invoked...]` prefix, the entire SKILL.md content (often 1K+ chars), and only then the user's actual instruction. Both `cli.py:6186` and `gateway/run.py:6031` pass this bloated `message` to `maybe_auto_title()`. The title generator truncates to 500 chars (`title_generator.py:29`), so it only sees skill boilerplate.

`run_conversation()` already computes `original_user_message` (line 6516) — the clean user input without skill injection. PR #4940 used it for memory providers with this exact comment:

Use original_user_message (clean input) — user_message may contain injected skill content that bloats / breaks provider queries.

But `original_user_message` was never added to the result dict, so callers couldn't access it.

Fix

1. run_agent.py: Added `original_user_message` to the result dict
2. cli.py: `result.get("original_user_message") or message`
3. gateway/run.py: Same pattern

Tests

4 tests: result dict includes original message, clean message preferred over bloated, fallback to raw message when missing, clean input fits within 500-char truncation.

```
4 passed
```

[dieutx]

Rebased this onto current main.\n\nLocal: pytest -q tests/test_auto_title_clean_message.py -> 4 passed

[teknium1]

Thanks for targeting a real title-quality problem: current main still passes the expanded message at cli.py:12558 and gateway/run.py:19030.

Problems

• The proposed original_user_message is not intrinsically clean. It is persist_user_message only when provided, otherwise the model-facing user_message (agent/turn_context.py:304; PR base run_agent.py:7925). Normal CLI skill invocation queues the expanded payload (cli.py:9037-9044) and only supplies that override for voice mode (cli.py:12319-12346), so the PR would still title from skill scaffolding.
• The added tests only construct local dicts and repeat the fallback expression; they never exercise either title caller with a skill-expanded message.
• The proposed result-dict hunk has moved: current result assembly is agent/turn_finalizer.py:399-428, following refactor 053025238434cfbf121873977b39888d7f27d1c1.

Suggested changes

• Use the existing extract_user_instruction_from_skill_message() (agent/skill_commands.py:58-112) at the title boundary and add CLI/gateway regression tests that assert maybe_auto_title receives the extracted instruction. Decide explicitly how a bare skill invocation should title, since the extractor returns None for that case.

Automated hermes-sweeper review.

[teknium1]

original_user_message is not guaranteed to be clean skill input: the turn setup defines it as persist_user_message only when supplied, otherwise the model-facing user_message (current agent/turn_context.py:304; PR base run_agent.py:7925). Normal CLI and gateway skill paths pass the expanded payload without that override, so this field would still contain the skill body. Use the canonical skill-instruction extractor before title generation instead.