backfill: continue on per-document errors instead of aborting

A single doc that isn't a real payslip (e.g., an RSU letter wrongly tagged
as payslip in Paperless) makes Claude return pay_date=null, which pydantic
rejects with ValidationError. Previously this killed the whole backfill at
the first bad doc, leaving 60 of 88 docs unprocessed.

Catch + log + continue so the backfill processes every doc. Failed docs
can be re-tagged or fixed individually later.

payslip_ingest/__main__.py

@@ -52,14 +52,24 @@ async def _backfill(tag: str, limit: int | None) -> None:
         bearer_token=os.environ["CLAUDE_AGENT_BEARER_TOKEN"],
     )
     processed = 0
+    failed = 0
     try:
         async for doc in paperless.list_tagged_documents(tag):
             if limit is not None and processed >= limit:
                 break
             doc_id = int(doc["id"])
-            result = await process_document(doc_id, session_factory, paperless, extractor)
+            try:
-            click.echo(f"doc_id={doc_id} status={result.status} validated={result.validated}")
+                result = await process_document(doc_id, session_factory, paperless, extractor)
+                click.echo(f"doc_id={doc_id} status={result.status} validated={result.validated}")
+            except Exception as exc:
+                # Don't let a single bad doc (wrong tag, non-payslip PDF, Claude
+                # hallucinating null fields) abort the whole backfill. Log + continue.
+                failed += 1
+                click.echo(f"doc_id={doc_id} status=failed error={type(exc).__name__}: {exc}",
+                           err=True)
+                log.exception("backfill: doc_id=%s failed", doc_id)
             processed += 1
+        click.echo(f"backfill complete: processed={processed} failed={failed}")
     finally:
         await paperless.aclose()
         await extractor.aclose()