Add the release diff for event replays and its issue packets

docs/uk_replay/RELEASE_REPLAY.md

@@ -275,7 +275,15 @@ Observed on 2.125.1 on 11 October (parameter tree only):
 
 `pipeline/diff_uk_event_runs.py --base <variant or root> --head <variant or root>`
 compares two complete namespaces, engine-free, after re-verifying each side's
-staging and comparison receipts. Keys:
+staging and comparison receipts against that side's receipt-bound registry
+snapshot. The current event registry does not replace a historical snapshot.
+Every inventoried artifact is checked by the staging and comparison validators,
+including artifacts that supply no numeric source counterpart. Each run
+manifest's execution identity, registry and artifact inventory, supported and
+unsupported years, and measure-year grids must agree with the verified staging
+rows and tally. Inertness comes from the staging manifest named and verified by
+comparison provenance. Inconsistent receipts are refused before a diff is
+written. Keys:
 
 * source row: `(event, source_row_id)`;
 * measure-year total: `(event, measure_key, year)`;
@@ -289,9 +297,9 @@ bin transitions.
 |---|---|
 | `numerical` | Both sides numeric and `abs(head − base) > max(£1,000,000, 0.01 × max(abs(head), abs(base)))`. Strictly greater. |
 | `structural` | Sign reversal or a zero crossing (zero means `abs(x) ≤ £0.50`, accounting roundoff); artifact schema or orientation changed; non-finite value. Regardless of size. |
-| `coverage` | A counterpart, artifact pair or source row present on one side only, in a year both sides support. |
-| `mapping` | A row's `head_variables`, an artifact's `head_channels` or `head_definitions`, or the registry SHA-256 differs. |
-| `support` | A year supported on one side only. |
+| `coverage` | A counterpart, artifact pair or source row present on one side only, in a year both sides support. The shared-support requirement also applies to rows outside household scope or without an expressible counterpart. |
+| `mapping` | A row's `head_variables`, an artifact's `head_channels` or `head_definitions`, or the registry SHA-256 differs. Registry changes flag the event's affected records with both digests. |
+| `support` | A year supported on one side only. Source rows that cannot carry a counterpart do not receive this flag. |
 | `inertness` | `inert_construction` on one side only, or identical worlds in an in-force year on one side only. |
 | `construction` | Executed `baseline_world`/`reform_world` reform dicts differ, or a premise check's result differs. A changed current-law value of a touched parameter is recorded as context on the row, not flagged: it is what a model release is expected to change. |
 
@@ -300,14 +308,24 @@ recorded once per diff; they explain moves and are not flags on rows.
 
 Outputs: `DIFF.json` (canonical, sorted), `DIFF.md`, and one issue packet per
 flagged `(event, measure, year)` under `packets/`. A packet holds the source
-cells (table, column, row number, source-row id, OBR value), both execution
+cells (table, column, row number, source-row id, OBR value) and measure details
+from each registry with base/head attribution, both execution
 identities, registry and artifact digests, the executed reform dicts, the
 affected heads with base/head values, focused rerun commands for each
-variant, the statutory sources the registry already cites, and empty
+variant using that side's verified registry path and variant selection, the
+statutory sources each registry already cites with side attribution, and empty
 `minimal_case` and `statutory_evidence` sections for the investigating lane.
+Removed measures retain their base title, source cells and citations.
 Packets are evidence for the coordinating session, which files upstream
 issues; this pipeline files none.
 
+Diff invariants are checked with deterministic cases and Hypothesis properties
+in `tests/test_uk_event_diff.py`: identical sides raise no flags; swapping sides
+preserves each record's flag set and mirrors its values; every source row in
+either side appears once at source-row level; every flagged record appears in
+exactly one packet, including support-only records; numerical moves flag only
+strictly above the stated threshold; and coverage and support do not overlap.
+
 ## Qualification requests and receipts
 
 See [QUALIFICATION_INTERFACE.md](QUALIFICATION_INTERFACE.md) for the request

pipeline/compare_uk_event.py

@@ -1321,8 +1321,14 @@ def render_summary(
     return "\n".join(lines)
 
 
-def load_verified_comparison(path: Path, *, artifact_root: Path = ROOT) -> list[dict]:
-    """Bind summary-only reads to current source, stage, axes and artifact bytes."""
+def load_verified_comparison(
+    path: Path, *, artifact_root: Path = ROOT, registry_path: Path | None = None
+) -> list[dict]:
+    """Verify receipts against the current registry or an explicit snapshot.
+
+    Historical readers may select the receipt-bound registry snapshot. The
+    default preserves summaries' binding to the current event registry.
+    """
     provenance_path = path.parent / "COMPARISON_PROVENANCE.json"
     if not provenance_path.exists():
         raise EventComparisonError("summary requires COMPARISON_PROVENANCE.json")
@@ -1354,7 +1360,9 @@ def load_verified_comparison(path: Path, *, artifact_root: Path = ROOT) -> list[
         raise EventComparisonError("comparison replay-grid receipt differs")
     if registry.get("event_slug") != path.parent.name:
         raise EventComparisonError("summary registry event differs")
-    canonical_registry = (
+    if registry_path is not None and registry_path.resolve() != paths["registry"]:
+        raise EventComparisonError("comparison registry snapshot path differs")
+    canonical_registry = registry_path or (
         artifact_root / "data" / "uk" / "events" / f"{path.parent.name}_measures.json"
     )
     if (