Drafting a commit message leaves the staged files' objects in the repository #34594

Area: apps/server

Steps to reproduce

1. In a repository, change or add a file (for example a large generated file or dataset).
2. Run the Commit action with an empty message, so T3 prepares the commit context
   (prepareCommitContext) to generate one.
3. Let the generation fail or stop it before the commit runs (for example, the text-generation
   provider is signed out or offline).
4. Run git cat-file -e $(git hash-object <file>).

Expected behavior

Preparing the context uses a temporary index, so without a commit the repository is unchanged.

Actual behavior

The object exists, unreachable. prepareCommitContext sets only GIT_INDEX_FILE to a temporary index and runs
git add -A, and git add writes every staged blob into the repository's object store. When the commit
follows, it needs those blobs anyway; when it does not, they stay as unreachable loose objects, and
each attempt after the file changes adds another copy. gc prunes them only when maintenance runs
and they are older than its expiry (two weeks by default). For large generated files or datasets
that is their compressed size, close to the file size for data that does not compress, and later
maintenance (a cruft repack) has to read all of it.

Impact

Minor bug or occasional failure

Version or commit

main at a02bc75