Clarify offset audit and refresh export manifest

README.md

@@ -42,7 +42,7 @@ The `related_mentions` config contains 2,384 positive spatial-linkage mentions u
 - `dataset` in passage rows is the **review collection/lane**, not the source corpus. `passage_origin` is the source prefix from the original passage key; `key` remains unchanged and preserves source document/page/chunk identity. `source_document_id` joins to `documents.document_id`.
 - `document_metadata` provides typed metadata. `documents` config has one row per reviewed document; `canonical_record_json` preserves the full canonical source record used by the analysis join, including every source field (such as source content/chunk/entity metadata). `metadata_source_file` identifies the canonical catalog file. Project id, title, document type, country, source URLs and all remaining fields are retained.
 - The original focus-sidecar fields (`h`, `key`, `dataset`, `split`, `focus`, `proposal_keep_only`, `passage`, `annotations`, validation/write timestamps) remain intact in passage records.
-- Original annotation character offsets are never rewritten. `related_mentions` adds `render_start`/`render_end`/`render_span` for five whitespace-only alignment recoveries; original offsets remain inside `annotation`.
+- Original annotation character offsets are never rewritten. Five source span slices do not exactly match passage text: four differ only by whitespace, and one end offset omits a closing parenthesis. The report's whitespace-flexible locator recovers all five; among related mentions, two need text recovery and one needs an end-offset adjustment. `related_mentions` adds passage-aligned `render_start`/`render_end`/`render_span`, preserving original span and offsets inside `annotation`.
 - `reports/pad-pid-geospatial-analysis.html` and the linked dataset census are refreshed from the same focus sidecar and document catalog.
 - `metadata/manifest.json` records source hashes, file hashes, split/class counts, and validation results.
 

metadata/manifest.json

@@ -54,7 +54,10 @@
       "refugee_pads": 579,
       "umar_pads": 2109
     },
-    "whitespace_alignment_adjustments": 1
+    "whitespace_alignment_adjustments": 1,
+    "source_span_text_mismatches": 5,
+    "related_span_text_mismatches": 2,
+    "non_whitespace_offset_mismatches": 1
   },
   "quality_checks": {
     "duplicate_focus_occurrences_skipped": 0,
@@ -64,12 +67,13 @@
     "full_canonical_document_record_preserved_as_json": true,
     "related_mentions_match_report_total": true,
     "passage_rows_roundtrip": true,
-    "related_rows_roundtrip": true
+    "related_rows_roundtrip": true,
+    "all_offset_mismatches_recovered_by_report_locator": true
   },
   "files": {
     "README.md": {
-      "sha256": "ea50e1bf800fd8175c0e96003c4ffd57115ea05720581c9cafec7503478a2af1",
-      "bytes": 4473
+      "sha256": "e114061c0e357cac0dc90d4ed0248081c5274656f2ef20a218d506704be29830",
+      "bytes": 4742
     },
     "data/documents.jsonl": {
       "sha256": "5fa57bc006762b1f7f049ab0baf42338ed83dfb58e13b366c9e0a4ba8dde7cf8",