Report what the forge said when a merge is refused

[ameyypawar]

🧠 Changes

merge_pull_request now reports what GitHub said, not just the status it said it with, by using response_error() — the helper the rest of the client already reaches for, including update_pull_request ten lines above it.

response_error() in turn surfaces GitHub's message rather than the whole JSON object, so a toast reads 405 Method Not Allowed: <sentence> instead of that sentence buried between documentation_url and status. Bodies carrying errors are passed through whole, since that array is the part of a validation failure worth reading and message is only its headline.

GitLab's merge_merge_request had the same gap, in a file that already spells {status} - {error_text} at five other sites. One line.

☕️ Reasoning

Merging with a method the repository has disabled fails today with nothing to act on. #8403 describes it as two problems stacked: the picker offers all three methods because nothing fetches allow_merge_commit / allow_squash_merge / allow_rebase_merge, and the rejection that follows is reduced to a status code. This is the second one. The first is UI work tracked separately in #8654, and greying options out is easier to judge once the failure is legible.

The change is small because the intent was already there: response_error()'s own doc comment says "so GitHub's error message reaches the user", four sites in stacks.rs use it, and #15212 is adding seven more. merge_pull_request was the one that missed it.

Three things this deliberately does not do:

• response_error() itself stays uncovered. Reaching it needs a real reqwest::Response, and there is no mock-HTTP anywhere in the workspace; extracting a one-line format! purely to test it seemed worse than the gap. What it delegates to is tested, fallbacks included.
• The five remaining response.status() sites in client.rs are untouched, keeping this a bug fix rather than a sweep. Happy to convert them in one go if you would prefer that.
• Error text now reaches the telemetry showError sends. Preferring message narrows what that carries compared with raw bodies, which is part of why errors is a carve-out rather than the general rule.

I have not reproduced a 405 against a repository with the method disabled, so the exact wording GitHub uses is taken from the issue rather than something I have seen. Nothing in the change depends on it.

🎫 Affected issues

Addresses: #8403