fix(client): explain missing Gemini thought signatures

[nightt5879]

Summary

Fixes #6048.

When a compatible gateway rejects a Gemini tool-call replay with Function call is missing a thought_signature in functionCall parts, Codewhale now explains how to recover: select the built-in google provider with its default endpoint, or a gateway that preserves signatures, then start a new session.

The existing shared HTTP error formatter applies this guidance to streaming and non-streaming requests only after an explicit missing-signature 400. It retains bounded provider details, preserves quota and HTML handling, and remains safe to format twice. Gateways that manage signatures themselves continue to accept unsigned history.

Thanks to @Hmbown for identifying the missing recovery guidance and the gateway compatibility constraint.

Testing

• cargo check -p codewhale-tui --lib --locked — passed.
• cargo fmt --all -- --check, git diff --check, provider registry and dead-code budget checks — passed.
• Version/changelog synchronization gate and 12 public-copy/changelog tests — passed; changelog edits are reserved for merge time per repository policy.
• Regression proof with the production change removed: 17 passed, 3 failed, including the real local HTTP request returning the original unhelpful 400.
• cargo test -p codewhale-tui --lib --all-features --locked -- thought_signature llm_client:: client::chat:: client::tests:: client::responses:: client::anthropic:: --test-threads=1 — 409 passed, 0 failed.
• Cloud Linux Clippy on Rust 1.98.1 passed. Local full-workspace Clippy on the same toolchain stops on three existing Windows-only warnings in child_env.rs, mcp.rs, and lib.rs; all three files match main. Three-platform CI tests are pending.

The full Linux web CI suite passed all 446 tests. The broader local web suite reached 393 passing tests, with unrelated Windows CRLF/POSIX-path failures. Local Rust tests run with the repository's isolated-home wrapper. Network cases run serially with ambient proxy variables removed from the test subprocess; the unrelated MCP network cases passed under that configuration.

Checklist

• Preserved report credit under Unreleased; addressed Devin feedback in d585f2d and resolved the review thread. Devin completed its latest-head analysis with no additional threads.
• Added reproduced-regression coverage, including successful gateway-managed replay.
• Used existing error handling rather than introducing a new layer.
• Full workspace Clippy and CI completed.

---

[Hmbown]

Thanks for the fix and the transport-level regression coverage. I harvested this into the local 0.9.13 release lane as 157c5e1ea7, preserving your authorship and contributor credit. It replaces the lane's earlier overlapping hint so there is one bounded, idempotent recovery path, including for streaming and non-streaming gateway errors.

The focused run passed all 71 selected tests, including your new transport and sanitization tests. The root npm and web gates also passed. The complete final-head sweep, live terminal verification, and hosted CI still need to finish before publication; this is not yet a released fix.

[github-actions]

Thanks @nightt5879 — your contribution landed in 157c5e1ea736 on main:

fix(client): explain Gemini signature recovery after gateway transport

Closing this PR now that the code is on main. Credit lives in the commit message and (where applicable) the CHANGELOG.md entry for the next release. Apologies for not closing this at the time of the merge — the auto-close workflow is new in v0.8.31.

If you want to land more work and would prefer your future PRs merge cleanly without a harvest step, the CONTRIBUTING.md doc has a short note on what makes a contribution mergeable as-is.