diff --git a/CLAUDE.md b/CLAUDE.md
index e627247..09a364f 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -92,8 +92,9 @@ does on one that was aborted. So a caller must also put `saved_staged_patch`
 back on the paths `apply_abort()` never sees — the `RebaseOutcome::Completed`
 arm and its `after_continue` handler — through
 `git::restore_staged_after_rebase`, which applies three-way and never fails its
-caller. `weave::run_rebase_or_abort` does it for its own callers (Specs 004 and
-014).
+caller, and on that arm goes before `transaction::delete` so a failed delete
+leaves the index as the user had it — with one exception (Spec 014).
+`weave::run_rebase_or_abort` does it for its own callers (Specs 004 and 014).
 
 A `weave::run_rebase_protecting` caller that can pause must also record its
 protected commits in `LoomState.protect`, as full object names; `loom continue`
diff --git a/specs/014-continue-abort.md b/specs/014-continue-abort.md
index a8d3df9..745c58a 100644
--- a/specs/014-continue-abort.md
+++ b/specs/014-continue-abort.md
@@ -481,14 +481,17 @@ entry either way.)
 
 A command that saves `saved_staged_patch` MUST therefore restore it on every
 exit: the `RebaseOutcome::Completed` arm, its `after_continue` handler, and the
-abort. The restore applies three-way, never resetting the index first: the patch
-is a HEAD-to-index diff, and a plain apply would refuse the whole of it over
-either a staged *new* file whose entry survived the autostash or a hunk whose
-context the rebase rewrote. Resetting to make it apply does not help and can
-lose work: on the success path HEAD has moved, so the reset lands the index on
-the *new* HEAD while the patch is against the old one — and it has already
-dropped what the autostash preserved, which the failing apply then cannot put
-back.
+abort. On that arm the restore MUST precede `transaction::delete`, so a delete
+that fails still leaves the index as the user had it; the exception is an arm
+whose own work reads the index to undo itself, which fold's uncommit does, and
+there the restore follows that work. The restore applies three-way, never
+resetting the index first: the patch is a HEAD-to-index diff, and a plain apply
+would refuse the whole of it over either a staged *new* file whose entry
+survived the autostash or a hunk whose context the rebase rewrote. Resetting to
+make it apply does not help and can lose work: on the success path HEAD has
+moved, so the reset lands the index on the *new* HEAD while the patch is
+against the old one — and it has already dropped what the autostash preserved,
+which the failing apply then cannot put back.
 
 The restore is best-effort and MUST NOT fail its caller: it runs after that
 command's own rewrite has landed, so an error here would report a rewrite that
diff --git a/src/absorb.rs b/src/absorb.rs
index 62201f9..a6fdbd5 100644
--- a/src/absorb.rs
+++ b/src/absorb.rs
@@ -302,10 +302,10 @@ fn apply_plan(repo: &Repository, workdir: &Path, git_dir: &Path, plan: AbsorbPla
         .map_err(|e| transaction::roll_back_failed_rebase(workdir, git_dir, &state, e))?;
     match outcome {
         RebaseOutcome::Completed => {
+            git::restore_staged_after_rebase(workdir, &saved_staged);
             transaction::delete(git_dir)?;
             post_absorb(
                 workdir,
-                &saved_staged,
                 skipped_patch.as_deref(),
                 plan.num_hunks,
                 plan.num_files,
@@ -422,9 +422,9 @@ pub fn after_continue(
 ) -> Result<()> {
     let ctx: AbsorbContext =
         serde_json::from_value(context.clone()).context("Failed to parse absorb resume context")?;
+    git::restore_staged_after_rebase(workdir, &rollback.saved_staged_patch);
     post_absorb(
         workdir,
-        &rollback.saved_staged_patch,
         ctx.skipped_patch.as_deref(),
         ctx.num_hunks,
         ctx.num_files,
@@ -432,17 +432,16 @@ pub fn after_continue(
     )
 }
 
-/// Post-rebase work: restore staged/skipped patches and print success message.
+/// Post-rebase work: re-apply the skipped patch and print the success message.
+/// The staged patch goes back before the state file does, so both callers
+/// restore it themselves.
 fn post_absorb(
     workdir: &Path,
-    saved_staged: &str,
     skipped_patch: Option<&str>,
     num_hunks: usize,
     num_files: usize,
     num_commits: usize,
 ) -> Result<()> {
-    git::restore_staged_after_rebase(workdir, saved_staged);
-
     if let Some(patch) = skipped_patch
         && let Err(e) = git::apply_patch(workdir, patch)
     {
diff --git a/src/commit.rs b/src/commit.rs
index f0f3ff8..ef11598 100644
--- a/src/commit.rs
+++ b/src/commit.rs
@@ -188,8 +188,8 @@ pub fn run(
         protect: vec![head_oid.to_string()],
     };
     transaction::save(&git_dir, &state)?;
-    // The state file owns the patch from here: `post_commit` puts it back on
-    // success, `Rollback` on abort.
+    // The state file owns the patch from here: the `Completed` arm puts it
+    // back on success, `Rollback` on abort.
     let saved_staged = staged_aside.release();
 
     let base = graph.base_oid.to_string();
@@ -200,8 +200,9 @@ pub fn run(
         .map_err(|e| transaction::roll_back_failed_rebase(&workdir, &git_dir, &state, e))?;
     match outcome {
         RebaseOutcome::Completed => {
+            git::restore_staged_after_rebase(&workdir, &saved_staged);
             transaction::delete(&git_dir)?;
-            post_commit(&workdir, &branch_name, &saved_staged)?;
+            post_commit(&workdir, &branch_name)?;
         }
         RebaseOutcome::Stopped => {
             transaction::warn_paused(&workdir, "commit");
@@ -225,13 +226,13 @@ pub fn after_continue(
     let saved_staged = ctx
         .saved_staged
         .unwrap_or_else(|| rollback.saved_staged_patch.clone());
-    post_commit(workdir, &ctx.branch_name, &saved_staged)
+    git::restore_staged_after_rebase(workdir, &saved_staged);
+    post_commit(workdir, &ctx.branch_name)
 }
 
-/// Post-rebase work: restore staged changes and print success message.
-fn post_commit(workdir: &Path, branch_name: &str, saved_staged: &str) -> Result<()> {
-    git::restore_staged_after_rebase(workdir, saved_staged);
-
+/// Post-rebase work: print the success message. The restore runs before the
+/// state file goes, so both callers do it themselves.
+fn post_commit(workdir: &Path, branch_name: &str) -> Result<()> {
     let new_hash = git::rev_parse(workdir, branch_name)?;
 
     msg::success(&format!(
diff --git a/src/fold.rs b/src/fold.rs
index 3999bb8..167319f 100644
--- a/src/fold.rs
+++ b/src/fold.rs
@@ -2575,6 +2575,10 @@ fn fold_commit_to_unstaged(repo: &Repository, commit_hash: &str) -> Result<()> {
             })?;
         let staged = match outcome {
             RebaseOutcome::Completed => {
+                // Not restore-then-delete as elsewhere (Spec 014): the apply
+                // below undoes itself by diffing the working tree against the
+                // index, so staged work put back first would change what it
+                // reads as its own.
                 transaction::delete(&git_dir)?;
                 if !diff.is_empty()
                     && let Err(e) = git::apply_patch_to_worktree(workdir, &diff)