[BUG] Desktop app: Edit approval card shows no diff unless Verbose transcript view is enabled (still present in 1.26832.0)

Summary

In the Claude desktop app, the Edit/Write approval card renders only the file name and a
line count (Update(path) / Added N lines) with no diff body. The red/green, line-numbered
diff appears only if Verbose transcript view is enabled.

This is the same regression as #48198, which was closed as not planned. Refiling because it is
still reproducible many app versions later (that report cited app 1.2581.0; this is 1.26832.0),
and because the verbose workaround has a side effect that makes it a poor substitute.

Environment

| Component | Version |
| --- | --- |
| Claude desktop app | 1.26832.0 (MSIX: Program Files\WindowsApps\Claude_1.26832.0.0_x64__pzs8sxrjxfjjc) |
| claude-code bundled in the app | 2.1.222 (%APPDATA%\Claude\claude-code\2.1.222\claude.exe) |
| claude-code standalone CLI | 2.1.223 (~\.local\bin\claude.exe) |
| OS | Windows 11 Enterprise 26200 |

Steps to reproduce

1. Open a project in the desktop app with the default (non-verbose) transcript view.
2. Ask Claude to modify a source file, so that it issues an Edit tool call.
3. Observe the permission card.

Expected: a red/green, line-numbered diff of the old and new text, as the terminal CLI renders
for the same tool call.

Actual: file name and "Added N lines" only — no diff body, nothing indicating what will change.

Workaround: switch the transcript view to Verbose. The diff then renders correctly.

Why the workaround isn't sufficient

Verbose is a single global toggle, so it also expands thinking blocks that are otherwise collapsed
to a one-line Cogitated for Ns summary. Reviewing diffs therefore requires permanently reading
full reasoning transcripts.

Per the settings documentation there is no way to keep verbose tool output while leaving thinking
collapsed: alwaysThinkingEnabled and MAX_THINKING_TOKENS disable extended thinking rather
than hide its display, which degrades output quality rather than just tidying the transcript.

Impact

Without the diff, approving an edit means approving blind. The alternative — asking the model to
paste the proposed change into chat as markdown — is not equivalent and not trustworthy: chat text
is composed separately from the tool call, so there is no guarantee it matches the bytes that will
be written, and it carries neither syntax highlighting nor real file line numbers. The permission
card is the only preview that is generated from the actual old_string/new_string of the call,
which is what makes it the review surface that matters.

In practice this turns a routine one-line fix into a long back-and-forth, and it silently weakens
the review step that the permission prompt exists to provide.

Request

Render the diff body in Edit/Write permission cards regardless of transcript verbosity — or, failing
that, add a setting that shows tool-call diffs without also expanding thinking blocks.