fix: review the change that was made, not the one that was proposed

[thedancingdeveloper]

Closes #132.

Executor passed the model's own diff text to the reviewer. SessionExecutor
already passed git diff HEAD. The plain one was the wrong way round.

This matters because the tolerance ladder's whole purpose is to rescue patches
whose text is malformed — so the text and the resulting change are routinely
different.

Evidence, from a live run

Model's text (what the reviewer used to see):

@@ -0,0 +1,4 @@
+def multiply(a, b):
+    return a * b

What actually landed:

@@ -1,3 +1,7 @@
+def multiply(a, b):
+    return a * b
+
+
 """A deliberately tiny module, so an agent's diff is easy to judge."""

The reviewer rejected the item because it "could not verify that existing
functions and tests were kept untouched, because both files are shown as being
created from empty". Correct about the text; false about the change. The
checks had already passed against the applied tree, including the pre-existing
tests — the harness had proof the file was intact and showed the reviewer
something else.

Live effect

Same three-item backlog, before and after:

• Before: 0/3. Both attempted items planned, implemented and passed checks,
  then failed review on the artefact.
• After: 1/3 completed end to end — plan → implement → apply → checks →
  review → commit — and the remaining two now fail on their merits: the
  reviewer can see that a rescued hunk landed above the file's existing
  imports. That is Bug: a zero-context hunk is applied at line 1, putting code above module docstrings while checks stay green #133, a real defect the proposed diff could not express.

Rejecting good work was the visible half. The invisible half is worse: a gate
told to catch a change that "claims more than it did" could not, because it
was reading the claim.

Tests

Two, both failing without the change: the reviewer's prompt must contain the
surviving context and must not contain the proposed @@ -0,0 header; and the
prompt must show where a rescued hunk landed relative to existing content.

548+ tests, ruff, mypy . green.