diff: EOF-append to flat YAML list yields zero role:"changed" fragments (change signal lost) + whole-file fragmentation

[nikolay-e]

Observed

On a diff whose sole change is three lines appended at EOF of a large flat
YAML list file, diffctx . --diff HEAD~1 emits a context output that:

1. Carries zero role: "changed" fragments. Not a single fragment in the
   output is labeled as the change — grep -c 'role:' output.yaml → 0. The
   "what changed" signal is entirely lost even though the diff is non-empty and
   valid.
2. Does not surface the changed lines at all. The change adds new lines
   786–788 (@@ -785,3 +785,6 @@); the highest line range anywhere in the
   output is 785-785. The appended lines never appear as their own fragment.
3. Explodes the whole ~790-line file into ~118 tiny definition fragments
   (lines: "4-6", "14-14", "17-17", … "785-785"), i.e. nearly the
   entire unchanged file is dumped as context.
4. Also pulls in 2 unrelated files that the diff never touched.

Total: fragment_count: 120, ~5.9k tokens, for a 3-line append.

Expected

• The appended lines (786–788) surfaced as a role: "changed" fragment — the
  change must always be represented and labeled.
• Context scoped to the changed region + tight related context, not a
  whole-file fragmentation of a flat data-list file.

Repro

Public repo, deterministic:

git clone https://github.com/elastic/elasticsearch
cd elasticsearch
git checkout 4fc110bec          # "Mute ... MultiClusterSpecIT ... #153046"
diffctx . --diff HEAD~1 -f yaml > out.yaml
git diff HEAD~1 --stat          # ground truth: muted-tests.yml | 3 +++
grep -c 'role:' out.yaml        # → 0  (expected: >= 1 role: "changed")
grep -oE 'lines: "[0-9]+' out.yaml | grep -oE '[0-9]+' | sort -n | tail -1
                                # → 785 (the changed 786-788 never surface)
grep -cE '^  - path:' out.yaml  # → 120 fragments for a 3-line append

The changed file is muted-tests.yml, a flat top-level YAML list
(- class: … / method: … / issue: … entries) ~790 lines long.

Analysis

Two distinct symptoms, likely two root causes:

• (A) Lost change signal — the EOF-append hunk maps to new lines 786–788,
  but no rendered fragment is in core_ids, so role: "changed" is never set.
  Either the changed-fragment for an at-EOF append in a non-code list file is
  never created, or it is dropped/absorbed during selection/merge. This is a
  correctness bug — worse than over-selection — since the actual answer to
  "what changed" is absent. Verified the field is emitted normally on other
  repos (e.g. numpy HEAD~1 yields 5 role: "changed" fragments), so this is
  a real omission, not a formatting artifact.
• (B) Whole-file fragmentation — a flat YAML list is split into ~118
  per-entry definition fragments and most of the file is selected as context.
  This over-selection facet overlaps diffctx over-selection: PPR/edge relevance filter pulls in unrelated files (whole-tree over-dump + co-located k8s manifests) #65 (whole-tree over-dump); tracking (A)
  here because the missing role: "changed" is a separate, sharper failure.

Env

• diffctx 1.10.2 (pipx-installed binary)
• macOS; ran perl -e 'alarm 200; exec @ARGV' diffctx . --diff HEAD~1 (no
  hang — completed in well under the cap)

Related: #65 (over-selection). This issue is primarily about (A).