fix(core): keep signed thinking ahead of tool calls on errored replay

[Enough1122]

Issue for this PR

Closes#38620

Type of change

• Bug fix

What does this PR do?

Replaying an errored assistant message stripped its reasoning provider metadata (signature / redactedData) while still replaying the same message's tool_use parts. Anthropic requires every thinking block to precede its sibling tool_use when thinking is enabled, so any interrupted-or-failed step in a thinking + tool session made the follow-up request fail with a 400.

The reuse gate now keeps signed/redacted reasoning blocks - with their metadata - when the producing model is the replay target, even if the message carries an error. Unsigned hosted-style reasoning (OpenAI itemId/encrypted content) keeps demoting to text as before, cross-model replay is unchanged, and the existing "lowers failed assistant reasoning to text" case still passes untouched since its fixture has no signature.

How did you verify your code works?

• new test in packages/core/test/session-runner-message.test.ts: errored same-model message with signed reasoning + a tool call lowers to reasoning(with anthropic signature) followed by tool-call, instead of demoting
• bun test ./test/session-runner-message.test.ts -> 23 pass / 0 fail
• bun run typecheck (packages/core): clean

Screenshots / recordings

Not a UI change.

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

[github-actions]

Thanks for your contribution!

This PR doesn't have a linked issue. All PRs must reference an existing issue.

Please:

1. Open an issue describing the bug/feature (if one doesn't exist)
2. Add Fixes #<number> or Closes #<number> to this PR description

See CONTRIBUTING.md for details.

[github-actions]

The following comment was made by an LLM, it may be inaccurate:

Found several related PRs that address similar issues with thinking/reasoning metadata on errored replays:

Potential Related PRs:

1. fix(core): keep reasoning provider metadata on errored turns for valid Anthropic replay #44054 - fix(core): keep reasoning provider metadata on errored turns for valid Anthropic replay

   • Directly related: handles reasoning metadata preservation on errored turns for Anthropic replay

2. fix(core): replay thinking safely and drop unsettled tool calls from errored messages #42735 - fix(core): replay thinking safely and drop unsettled tool calls from errored messages

   • Related: addresses thinking replay safety and tool call handling from errored messages

3. fix(core): drop tool parts when a failed turn loses its reasoning block #40157 - fix(core): drop tool parts when a failed turn loses its reasoning block

   • Related: handles tool parts when reasoning blocks are lost during failed turns

4. fix(core): preserve reasoning metadata on errored assistant turns #38763 - fix(core): preserve reasoning metadata on errored assistant turns

   • Related: preserves reasoning metadata on errored assistant turns (issue V2: errored-message replay can 400 with Anthropic thinking + tool use #38620 area)

5. fix(opencode): drop truncated reasoning from replayed history #40148 - fix(opencode): drop truncated reasoning from replayed history

   • Related: handles reasoning blocks in replayed history

These PRs address overlapping concerns around preserving thinking/reasoning signatures and handling tool calls when messages have errors. PR #44054 appears most closely related as it also handles "reasoning provider metadata on errored turns for valid Anthropic replay."

[Enough1122]

Closing this in favor of the four earlier PRs covering the same ground - #44054, #42735, #40157, #38763 - which the maintainers can now pick between without a fifth candidate muddying the choice. If whichever lands ends up missing regression coverage, the test added here (errored same-model message with signed reasoning + tool call keeps its thinking block ahead of tool-use) should drop in cleanly.