fix(diff): read the target commit's blob, not the working tree

[r0h1tb]

Description

Both git-diff paths read the old side from the git object store via
_read_blob and the new side with open(file_path) — the working tree.

compute_diff_for_commits documents "added_nodes: New AST nodes in
to_commit", and ast-rag update takes --from-commit and --to-commit as
required options. So any pair other than <x>..HEAD on a clean tree parses
the current source and stamps it with to_commit.

update_from_git has the same line and that one writes to Neo4j — the graph
ends up holding code that never existed at the commit it is labelled with.
Uncommitted work already has its own path in get_workspace_diff, so the two
were not meant to be the same thing.

The new side now comes from _read_blob at the target commit. Absence there,
rather than absence on disk, decides the deletion branch — which also fixes the
case where a file is deleted at to_commit but still present locally.

When to_commit is HEAD and the tree is clean the blob and the file are
identical, so the common path is unchanged.

Related Issue

None filed — found while auditing the temporal/versioning claims in #59.

Type of Change

• Bug fix (non-breaking change that fixes an issue)
• New feature
• Breaking change
• Documentation update

Checklist

• My code follows the code style of this project
• I have added tests that prove my fix works
• All new and existing tests passed
• I have updated the documentation accordingly
• I have run ast-rag evaluate --all — needs live services; no query or scoring path touched
• My changes generate no new warnings

Testing

New tests/test_diff_reads_the_named_commit.py builds a throwaway git repo
where alpha sits on line 1 at A, line 2 at B, and line 7 in an uncommitted
edit, then asks for A..B. No Neo4j, Qdrant or network needed.

Without the fix all three fail, reporting the working tree:

AssertionError: `alpha` spans lines 2-3 at to_commit; got 7-8
 +  where 8 = ASTNode(... name='alpha', ... valid_from='b3817cb3ed9c...')

That valid_from is the point: the node is stamped with a commit whose source
it did not come from.

One of the three asserts the weaker property directly — editing an uncommitted
file must not change the reported diff of two fixed commits.

Suite: 257 passed / 1 skipped / 1 xfailed before, 260 / 1 / 1 after (3 new).

Note for review

update_from_git filters candidates through file_changed_since_last_index,
which hashes the file on disk. That is a separate pre-existing issue —
during a historical replay it can skip files whose current contents happen to
match the cache — and is left alone here.