fix: repair orphaned toolUse in the last session message during restore

[Zelys-DFKH]

Description

_fix_broken_tool_use explicitly skipped the last message in the session history,
relying on the agent-class fallback (_has_tool_use_in_latest_message) to handle
it. That fallback only works when the same agent instance re-enters the event loop
within the same process. When a different process restores a session that ended with
an orphaned toolUse — for example, after a runtime timeout while waiting for a
long-running tool — and a new user message is then appended, the orphaned toolUse
reaches the model without a corresponding toolResult, producing a
ValidationException.

The fix removes the guard and handles the last-message case the same way as interior
messages: append a synthetic toolResult with status: "error". The original tool
execution context is already lost at restore time, so reporting an error to the model
and letting it decide how to proceed is the correct behavior.

Related Issues

Resolves #2025

Documentation PR

N/A

Type of Change

Bug fix

Testing

• Updated test_fix_broken_tool_use_ignores_last_message (which asserted the old,
  incorrect behavior) to test_fix_broken_tool_use_repairs_orphaned_last_message,
  verifying that a synthetic toolResult is appended.
• I ran hatch run prepare

Checklist

• I have read the CONTRIBUTING document
• I have added any necessary tests that prove my fix is effective or my feature works
• I have updated the documentation accordingly
• I have added an appropriate example to the documentation to outline the feature, or no new docs are needed
• My changes generate no new warnings
• Any dependent changes have been merged and published

---

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[konippi]

This PR duplicates #2026, which has been open since April 1st — four days before this one was created. I'd also like to flag a broader concern: @Zelys-DFKH opened 21 PRs (#2056–#2077) against this repository within roughly two days (April 4–5), spanning unrelated areas such as streaming, Bedrock, MCP, multiagent, session management, hooks, Anthropic, and OpenTelemetry. The pattern — commenting on an issue and then opening a PR minutes later, across a wide range of topics — strongly suggests automated bulk PR generation, likely using an AI coding agent.

[Zelys-DFKH]

I've been doing thousands of annual commits like a madman since before AI was around. Mostly privately. I apologize if I took something you wanted: there was a lot of low-hanging fruit that gave me an itch.

On the duplicate: the search I ran didn't catch #2026 because its title has a typo, "toolUser" instead of "toolUse." Happy to close mine or port the description over to that one. Maintainers' call.

Your comment identifying AI writing patterns has three em dashes, double-hedged conclusions ("strongly suggests... likely using"), an eight-item list, the phrase "broader concern" as vague framing, and a subordinate clause inside a subordinate clause. That's as AI as it gets. Just worth noticing, as you said.

Either way, good catch on #2026. I missed it and would not have duped it but for its typo.