diff --git a/application/alembic/env.py b/application/alembic/env.py
index 68eadc01..08aae816 100644
--- a/application/alembic/env.py
+++ b/application/alembic/env.py
@@ -28,7 +28,13 @@ if settings.POSTGRES_URI:
     config.set_main_option("sqlalchemy.url", settings.POSTGRES_URI)
 
 if config.config_file_name is not None:
-    fileConfig(config.config_file_name)
+    # ``disable_existing_loggers`` defaults to True, which would switch off
+    # every ``application.*`` logger already imported by the time migrations
+    # run. ``app.py`` calls ``setup_logging()`` (line 16) BEFORE
+    # ``ensure_database_ready()`` (line 58), so the web tier never
+    # re-enables them and loses all application logging for the life of the
+    # process — on exactly the boot where a schema upgrade happened.
+    fileConfig(config.config_file_name, disable_existing_loggers=False)
 
 
 def run_migrations_offline() -> None:
diff --git a/application/api/answer/routes/answer.py b/application/api/answer/routes/answer.py
index 3fab4de2..4b5418d3 100644
--- a/application/api/answer/routes/answer.py
+++ b/application/api/answer/routes/answer.py
@@ -8,6 +8,9 @@ from application.api import api
 
 from application.api.answer.routes.base import answer_ns, BaseAnswerResource
 
+from application.api.answer.services.continuation_service import (
+    ResumeInProgressError,
+)
 from application.api.answer.services.persistence_policy import resolve_persistence
 from application.api.answer.services.stream_processor import StreamProcessor
 
@@ -164,6 +167,20 @@ class AnswerResource(Resource, BaseAnswerResource):
             extra_info = stream_result.get("extra")
             if extra_info:
                 result.update(extra_info)
+        except ResumeInProgressError as e:
+            # Another request already owns this conversation's continuation
+            # claim. Same contract as ``/stream`` and ``/v1/chat/completions``:
+            # an expected concurrency outcome the caller can retry, not a
+            # server fault, so it must not land in the 500 handler below and
+            # fill the ERROR channel with tracebacks.
+            logger.warning(
+                "/api/answer - resume already in progress for conversation %s",
+                data.get("conversation_id"),
+                extra={"error": str(e)},
+            )
+            return make_response(
+                {"error": str(e), "code": "resume_in_progress"}, 409,
+            )
         except Exception as e:
             logger.error(
                 f"/api/answer - error: {str(e)} - traceback: {traceback.format_exc()}",
diff --git a/application/api/answer/routes/base.py b/application/api/answer/routes/base.py
index 38791923..b591c947 100644
--- a/application/api/answer/routes/base.py
+++ b/application/api/answer/routes/base.py
@@ -1054,9 +1054,27 @@ class BaseAnswerResource:
                         # that sets it lives in the superseding request's own
                         # process. That is routine and must not page anyone.
                         # Anything else is a genuinely orphaned answer.
-                        superseded = self.conversation_service.was_superseded(
-                            reserved_message_id
-                        )
+                        # Guarded: this opens its own DB connection, and an
+                        # unguarded raise here would escape to the generic
+                        # ``except`` below — which yields a terminal ``error``
+                        # frame, re-finalizes, and releases the resume claim,
+                        # all for a row that is already gone. It would also
+                        # swallow the very alert this branch exists to raise.
+                        # A missing ``superseded_messages`` table (code
+                        # deployed ahead of ``alembic upgrade head``) lands
+                        # here too, so the fallback is the pre-tombstone
+                        # behaviour rather than a user-visible failure.
+                        try:
+                            superseded = self.conversation_service.was_superseded(
+                                reserved_message_id
+                            )
+                        except Exception:
+                            logger.exception(
+                                "was_superseded lookup failed for %s; "
+                                "treating the row as orphaned",
+                                reserved_message_id,
+                            )
+                            superseded = False
                         if superseded:
                             logger.info(
                                 "stream superseded: message row %s was replaced "
@@ -1415,9 +1433,10 @@ class BaseAnswerResource:
             # invisible, and every retry inside the window gets another 409.
             # Not on the ``StreamSuperseded`` path above — there the turn was
             # deliberately replaced and the row is legitimately gone.
+            claim_released = False
             if _continuation and conversation_id:
                 try:
-                    ContinuationService().release_claim(
+                    claim_released = ContinuationService().release_claim(
                         str(conversation_id),
                         decoded_token.get("sub", "local"),
                     )
@@ -1428,6 +1447,21 @@ class BaseAnswerResource:
                         exc_info=True,
                     )
             if reserved_message_id is not None:
+                # Releasing the claim above says "this turn is retryable"; the
+                # row must not simultaneously be stamped terminally failed in a
+                # way the retry cannot overwrite. A resume reuses the SAME
+                # ``reserved_message_id`` (stream_processor stores it in the
+                # persisted ``agent_config``), so the retry's own
+                # ``finalize_message(status="complete")`` is gated by
+                # ``only_if_non_terminal`` and lands ``ALREADY_FAILED`` — the
+                # user watches a correct answer stream in and finds
+                # "Response was terminated prior to completion" on reload.
+                # Mark the failure retryable so the reclaim hole in
+                # ``update_message_by_id`` lets that second answer through,
+                # exactly as it already does for the reconciler's own marker.
+                failure_metadata = dict(query_metadata or {})
+                if claim_released:
+                    failure_metadata["resume_retryable"] = True
                 try:
                     self.conversation_service.finalize_message(
                         reserved_message_id,
@@ -1436,7 +1470,7 @@ class BaseAnswerResource:
                         sources=source_log_docs,
                         tool_calls=tool_calls,
                         model_id=model_id or self.default_model_id,
-                        metadata=query_metadata if query_metadata else None,
+                        metadata=failure_metadata or None,
                         status="failed",
                         error=e,
                     )
diff --git a/application/api/user/reconciliation.py b/application/api/user/reconciliation.py
index 4415fdd5..6dff0205 100644
--- a/application/api/user/reconciliation.py
+++ b/application/api/user/reconciliation.py
@@ -30,6 +30,21 @@ logger = logging.getLogger(__name__)
 MAX_MESSAGE_RECONCILE_ATTEMPTS = 3
 
 
+def zero_summary() -> Dict[str, int]:
+    """The counter keys every tick reports, all at zero.
+
+    Single source of truth so the early-return and the caller's error
+    fallback cannot drift from the sweeps that actually populate them.
+    """
+    return {
+        "messages_failed": 0,
+        "tool_calls_failed": 0,
+        "ingests_stalled": 0,
+        "idempotency_pending_failed": 0,
+        "schedule_runs_failed": 0,
+    }
+
+
 def run_reconciliation() -> Dict[str, Any]:
     """Single tick of the reconciler. Five sweeps, FOR UPDATE SKIP LOCKED.
 
@@ -42,20 +57,10 @@ def run_reconciliation() -> Dict[str, Any]:
     sitting in ``pending`` until 24 h TTL.
     """
     if not settings.POSTGRES_URI:
-        return {
-            "messages_failed": 0,
-            "tool_calls_failed": 0,
-            "skipped": "POSTGRES_URI not set",
-        }
+        return {**zero_summary(), "skipped": "POSTGRES_URI not set"}
 
     engine = get_engine()
-    summary = {
-        "messages_failed": 0,
-        "tool_calls_failed": 0,
-        "ingests_stalled": 0,
-        "idempotency_pending_failed": 0,
-        "schedule_runs_failed": 0,
-    }
+    summary: Dict[str, Any] = zero_summary()
     # User-facing events to fan out once their DB writes have committed
     # (publish-after-commit). Each item is
     # ``(user_id, event_type, payload, scope)``.
diff --git a/application/api/user/tasks.py b/application/api/user/tasks.py
index b8974226..2cf4e545 100644
--- a/application/api/user/tasks.py
+++ b/application/api/user/tasks.py
@@ -60,15 +60,40 @@ logger = logging.getLogger(__name__)
 # ``retry_kwargs["countdown"]`` into it on every retry, so one shared dict
 # spread across all the decorators below would let concurrent retries race
 # and would permanently mutate this module constant.
+#
+# ``dont_autoretry_for`` is the necessary counterweight to that widened
+# envelope. ``autoretry_for=(Exception,)`` retries EVERYTHING, but a
+# ``DocumentParseError`` (unparseable, empty, or image-only file) fails
+# identically on every attempt — with factor 60 it now burns a median 3.5 min
+# and up to ~7 before reaching a terminal state, during which anything polling
+# ``/api/task_status`` (the wiki-convert and GraphRAG-enable modals) just keeps
+# reporting "pending". It is the DEFAULT here so a new durable task cannot
+# forget it; tasks needing a wider set pass their own via ``durable_task()``.
 DURABLE_TASK = dict(
     bind=True,
     acks_late=True,
     autoretry_for=(Exception,),
+    dont_autoretry_for=(DocumentParseError,),
     max_retries=3,
     retry_backoff=60,
 )
 
 
+def durable_task(**overrides) -> Dict:
+    """Return ``DURABLE_TASK`` with per-task overrides applied.
+
+    Needed because ``@celery.task(**DURABLE_TASK, dont_autoretry_for=...)``
+    would be a duplicate keyword argument.
+
+    Args:
+        **overrides: Task options replacing the shared defaults.
+
+    Returns:
+        The merged options dict, for ``@celery.task(**durable_task(...))``.
+    """
+    return {**DURABLE_TASK, **overrides}
+
+
 # operation tag for the poison-path source.ingest.failed event, per task.
 _INGEST_POISON_OPERATION = {
     "ingest": "upload",
@@ -103,10 +128,11 @@ def _emit_ingest_poison_event(task_name, bound):
     )
 
 
-# ``dont_autoretry_for``: a file that cannot be converted to text fails
-# identically on every attempt, so retrying only multiplies the log noise
-# before the same failure — go straight to the poison/failure path.
-@celery.task(**DURABLE_TASK, dont_autoretry_for=(DocumentParseError,))
+# ``dont_autoretry_for=(DocumentParseError,)`` now comes from ``DURABLE_TASK``:
+# a file that cannot be converted to text fails identically on every attempt,
+# so retrying only multiplies the log noise before the same failure — go
+# straight to the poison/failure path.
+@celery.task(**DURABLE_TASK)
 @with_idempotency(task_name="ingest", on_poison=_emit_ingest_poison_event)
 def ingest(
     self,
@@ -285,8 +311,9 @@ def _emit_attachment_poison_event(task_name, bound):
 # deterministic — retrying re-fails identically and multiplies log noise, so it
 # goes straight to the failure path.
 @celery.task(
-    **DURABLE_TASK,
-    dont_autoretry_for=(DataError, AttachmentRejectedError, DocumentParseError),
+    **durable_task(
+        dont_autoretry_for=(DataError, AttachmentRejectedError, DocumentParseError),
+    )
 )
 @with_idempotency(
     task_name="store_attachment", on_poison=_emit_attachment_poison_event,
@@ -722,7 +749,7 @@ def reconciliation_task(self):
     reconciler findings — at 30 s cadence a multi-minute outage would otherwise
     bury genuine alerts under dozens of identical stack traces.
     """
-    from application.api.user.reconciliation import run_reconciliation
+    from application.api.user.reconciliation import run_reconciliation, zero_summary
 
     try:
         return run_reconciliation()
@@ -733,14 +760,9 @@ def reconciliation_task(self):
         # SUCCESS because the exception is swallowed here, so this log is the
         # only channel that carries the stack.
         logger.exception("reconciliation_task failed; the next beat retries")
-        return {
-            "messages_failed": 0,
-            "tool_calls_failed": 0,
-            "ingests_stalled": 0,
-            "attachments_stalled": 0,
-            "schedule_runs_failed": 0,
-            "error": True,
-        }
+        # Keys come from ``zero_summary`` so a failed tick reports the same
+        # shape as a successful one; hand-writing them drifted immediately.
+        return {**zero_summary(), "error": True}
 
 
 @celery.task(bind=True, acks_late=False)
diff --git a/application/parser/embedding_pipeline.py b/application/parser/embedding_pipeline.py
index c850518d..b6d3a0a6 100755
--- a/application/parser/embedding_pipeline.py
+++ b/application/parser/embedding_pipeline.py
@@ -5,6 +5,7 @@ from retry import retry
 from tqdm import tqdm
 from application.core.settings import settings
 from application.events.publisher import publish_user_event
+from application.parser.file.base_parser import DocumentParseError
 from application.storage.db.repositories.ingest_chunk_progress import (
     IngestChunkProgressRepository,
 )
@@ -204,7 +205,12 @@ def embed_and_store_documents(
         if str(getattr(d, "text", getattr(d, "page_content", d)) or "").strip()
     ]
     if not docs:
-        raise ValueError(
+        # ``DocumentParseError``, not ``ValueError``: an empty or image-only
+        # file yields nothing on every attempt, so this must reach the Celery
+        # tasks' ``dont_autoretry_for`` and fail once. As a bare ``ValueError``
+        # it was swept up by ``autoretry_for=(Exception,)`` and re-failed
+        # identically across the whole backoff envelope.
+        raise DocumentParseError(
             "No text could be extracted from this file. It may be empty, "
             "image-only, or in an unsupported format."
         )
diff --git a/application/storage/db/repositories/conversations.py b/application/storage/db/repositories/conversations.py
index 6926f274..cd27bbea 100644
--- a/application/storage/db/repositories/conversations.py
+++ b/application/storage/db/repositories/conversations.py
@@ -896,16 +896,23 @@ class ConversationsRepository:
         True, the update is gated so a late finalize cannot retract a
         reconciler-set ``failed`` (or a prior ``complete``).
 
-        ``reclaim_reconciler_failed`` punches one hole in that gate: a row
-        the **reconciler** failed for staleness may still be finalized by the
-        stream that owns it, because such a stream demonstrably was not dead.
-        Production showed a 20-minute tool loop swept at minute 6 and then
-        completing normally at minute 20, its answer refused and replaced by
-        "Response was terminated prior to completion". The hole is deliberately
-        narrow — it matches only the reconciler's own error string, so genuine
-        terminal failures (client disconnect, provider errors, the pending-tool
-        cleanup task) stay terminal, and it skips rows whose awaiting-approval
-        prompt the reconciler already revoked.
+        ``reclaim_reconciler_failed`` punches two narrow holes in that gate,
+        both for rows whose ``failed`` stamp a later successful answer
+        disproves:
+
+        1. A row the **reconciler** failed for staleness. Such a stream
+           demonstrably was not dead — production showed a 20-minute tool loop
+           swept at minute 6 and then completing normally at minute 20, its
+           answer refused and replaced by "Response was terminated prior to
+           completion". Matched on the reconciler's own error string, and
+           skipping rows whose awaiting-approval prompt it already revoked.
+        2. A row marked ``resume_retryable`` by the stream handler, which sets
+           it only when it also released the continuation claim — i.e. it
+           explicitly handed the turn back for a retry that reuses this same
+           row id.
+
+        Everything else stays terminal: genuine failures (client disconnect,
+        provider errors, the pending-tool cleanup task) are unaffected.
 
         Reclaiming also strips the reconciler's bookkeeping keys, which
         otherwise survive the metadata merge and leave a ``complete`` row
@@ -922,7 +929,7 @@ class ConversationsRepository:
             fields: Field/value pairs to write.
             only_if_non_terminal: Gate the write on a non-terminal status.
             reclaim_reconciler_failed: Also allow rows the reconciler failed
-                for staleness.
+                for staleness, and rows a failed resume marked retryable.
 
         Returns:
             MessageUpdateOutcome: What happened to the row.
@@ -946,6 +953,7 @@ class ConversationsRepository:
         # does this fire" query.
         reclaim_strip = (
             " - 'error' - 'reconcile_attempts' - 'last_reconcile_seen_heartbeat'"
+            " - 'resume_retryable'"
             if reclaim_reconciler_failed
             else ""
         )
@@ -1002,6 +1010,18 @@ class ConversationsRepository:
                     "AND COALESCE("
                     "(message_metadata->>'reconciler_cleared_approval')::boolean,"
                     " false) = false"
+                    ") OR ("
+                    # Second hole, same principle: the stream handler that
+                    # failed a resume also handed the turn back by releasing
+                    # the continuation claim, so it stamped the row
+                    # ``resume_retryable``. The retry reuses this very row
+                    # (the reserved id lives in the persisted agent_config);
+                    # without this it would land ALREADY_FAILED and the
+                    # user's second, successful answer would be discarded.
+                    "status = 'failed' "
+                    "AND COALESCE("
+                    "(message_metadata->>'resume_retryable')::boolean,"
+                    " false) = true"
                     "))"
                 )
             update_where.append(terminal_gate)
diff --git a/frontend/src/conversation/AnswerFlow.tsx b/frontend/src/conversation/AnswerFlow.tsx
index dd506064..c9f32892 100644
--- a/frontend/src/conversation/AnswerFlow.tsx
+++ b/frontend/src/conversation/AnswerFlow.tsx
@@ -29,8 +29,10 @@ type AnswerFlowProps = {
   agentId?: string;
   /** Set when the bubble already carries its own progress UI (a research run). */
   suppressStatusLine?: boolean;
-  /** Artifacts produced on this turn, so ``sandbox:`` links can reach them. */
+  /** Every artifact in the conversation, so ``sandbox:`` links can reach them. */
   artifacts?: SandboxArtifact[];
+  /** This turn's own artifacts, which win when a link names a bare filename. */
+  turnArtifacts?: SandboxArtifact[];
   onOpenArtifact?: (artifact: { id: string; toolName: string }) => void;
   renderApproval: (toolCall: ToolCallsType) => React.ReactNode;
   renderWikiWrite: (
@@ -52,6 +54,7 @@ export default function AnswerFlow({
   agentId,
   suppressStatusLine,
   artifacts,
+  turnArtifacts,
   onOpenArtifact,
   renderApproval,
   renderWikiWrite,
@@ -131,6 +134,7 @@ export default function AnswerFlow({
               content={message}
               isStreaming={isStreaming}
               artifacts={artifacts}
+              turnArtifacts={turnArtifacts}
               onOpenArtifact={onOpenArtifact}
             />
           </div>
diff --git a/frontend/src/conversation/ConversationBubble.tsx b/frontend/src/conversation/ConversationBubble.tsx
index 9328b39f..05b14000 100644
--- a/frontend/src/conversation/ConversationBubble.tsx
+++ b/frontend/src/conversation/ConversationBubble.tsx
@@ -439,6 +439,10 @@ const ConversationBubble = forwardRef<
             // would be a second live indicator away from the point of action.
             suppressStatusLine={Boolean(research)}
             artifacts={conversationArtifacts ?? completedArtifacts}
+            // The chip rail below renders from this turn's artifacts alone, so
+            // a bare-filename link has to resolve against the same set or the
+            // two controls in one bubble open different files.
+            turnArtifacts={completedArtifacts}
             onOpenArtifact={onOpenArtifact}
             renderApproval={(toolCall: ToolCallsType) => (
               <div className="fade-in mt-4 mr-5 ml-6">
diff --git a/frontend/src/conversation/MarkdownAnswer.tsx b/frontend/src/conversation/MarkdownAnswer.tsx
index cafb1274..903b7a98 100644
--- a/frontend/src/conversation/MarkdownAnswer.tsx
+++ b/frontend/src/conversation/MarkdownAnswer.tsx
@@ -70,12 +70,19 @@ export default function MarkdownAnswer({
   content,
   isStreaming,
   artifacts,
+  turnArtifacts,
   onOpenArtifact,
 }: {
   content: string;
   isStreaming?: boolean;
-  /** Artifacts produced on this turn, in creation order (``A1`` is the first). */
+  /**
+   * Every artifact in the conversation, in creation order (``A1`` is the
+   * first) — refs are conversation-scoped, so a link may name an earlier turn's
+   * file.
+   */
   artifacts?: SandboxArtifact[];
+  /** This turn's own artifacts; the filename fallback prefers them. */
+  turnArtifacts?: SandboxArtifact[];
   onOpenArtifact?: (artifact: { id: string; toolName: string }) => void;
 }) {
   const [isDarkTheme] = useDarkTheme();
@@ -132,7 +139,11 @@ export default function MarkdownAnswer({
                   // chip, but the model links it with a `sandbox:`/`artifact:`
                   // URL no browser can open. Point the link at the chip
                   // instead, and never leave a dead anchor behind.
-                  const sandboxLink = resolveSandboxLink(href, artifacts);
+                  const sandboxLink = resolveSandboxLink(
+                    href,
+                    artifacts,
+                    turnArtifacts,
+                  );
                   if (sandboxLink.kind === 'plain') {
                     return <>{children}</>;
                   }
@@ -185,7 +196,11 @@ export default function MarkdownAnswer({
                   // scheme survives and renders a broken-image box beside the
                   // chip that opens the very same file.
                   const source = typeof src === 'string' ? src : undefined;
-                  const sandboxLink = resolveSandboxLink(source, artifacts);
+                  const sandboxLink = resolveSandboxLink(
+                    source,
+                    artifacts,
+                    turnArtifacts,
+                  );
                   if (sandboxLink.kind === 'artifact') {
                     const { artifact } = sandboxLink;
                     return renderArtifactChip(
diff --git a/frontend/src/conversation/sandboxLinks.test.ts b/frontend/src/conversation/sandboxLinks.test.ts
index 99ca02d6..e1569359 100644
--- a/frontend/src/conversation/sandboxLinks.test.ts
+++ b/frontend/src/conversation/sandboxLinks.test.ts
@@ -296,3 +296,63 @@ describe('scheme-less and file: hrefs', () => {
     });
   });
 });
+
+// Two turns each produced a file called `report.pdf`. That is not exotic:
+// `_filename()` in artifact_generator.py hands EVERY html artifact the name
+// `report.html`, and re-running a script that writes `chart.png` collides too.
+describe('a filename that two turns both produced', () => {
+  const turnOne: SandboxArtifact = {
+    id: 'id-old',
+    label: 'report.pdf',
+    ref: 'A1',
+    toolName: 'artifact_generator',
+  };
+  const turnTwo: SandboxArtifact = {
+    id: 'id-new',
+    label: 'report.pdf',
+    ref: 'A2',
+    toolName: 'artifact_generator',
+  };
+  // What ConversationMessages passes: the whole conversation, in order.
+  const conversation = [turnOne, turnTwo];
+
+  it("resolves turn one's link to turn one's file, not the newest copy", () => {
+    // Without the turn scope this returned `id-new`, so the prose link opened
+    // a different file from the download chip in the very same bubble.
+    expect(
+      resolveSandboxLink('sandbox:/mnt/data/report.pdf', conversation, [
+        turnOne,
+      ]),
+    ).toEqual({ kind: 'artifact', artifact: turnOne });
+  });
+
+  it("resolves turn two's link to turn two's file", () => {
+    expect(
+      resolveSandboxLink('sandbox:/mnt/data/report.pdf', conversation, [
+        turnTwo,
+      ]),
+    ).toEqual({ kind: 'artifact', artifact: turnTwo });
+  });
+
+  it('still reaches an earlier turn when this turn produced nothing', () => {
+    // The fallback that keeps "the CSV from earlier" working.
+    expect(
+      resolveSandboxLink('sandbox:/mnt/data/report.pdf', conversation, []),
+    ).toEqual({ kind: 'artifact', artifact: turnTwo });
+  });
+
+  it('matches a ref conversation-wide even when this turn owns a file', () => {
+    // Refs are unambiguous, so they must NOT be narrowed — this is the
+    // cross-turn resolution the conversation-wide list was added for.
+    expect(resolveSandboxLink('artifact:A1', conversation, [turnTwo])).toEqual({
+      kind: 'artifact',
+      artifact: turnOne,
+    });
+  });
+
+  it('matches an id conversation-wide even when this turn owns a file', () => {
+    expect(
+      resolveSandboxLink('sandbox:/artifact/id-old', conversation, [turnTwo]),
+    ).toEqual({ kind: 'artifact', artifact: turnOne });
+  });
+});
diff --git a/frontend/src/conversation/sandboxLinks.ts b/frontend/src/conversation/sandboxLinks.ts
index dafc0cb7..2e315810 100644
--- a/frontend/src/conversation/sandboxLinks.ts
+++ b/frontend/src/conversation/sandboxLinks.ts
@@ -95,9 +95,9 @@ const SHORT_REF = /^a(\d+)$/i;
 const ANY_SCHEME = /^[a-z][a-z0-9+.-]*:/i;
 
 /**
- * Last match wins. Within one turn the model's prose means the file as it
- * stands after the *last* write, so a rerun that overwrites ``report.pdf``
- * should open the new copy rather than the one it replaced.
+ * Last match wins: the model's prose means the file as it stands after the
+ * *last* write, so a rerun that overwrites ``report.pdf`` should open the new
+ * copy rather than the one it replaced.
  */
 function findLast<T>(
   items: T[],
@@ -122,8 +122,11 @@ function decodeSegment(segment: string): string {
  *
  * Args:
  *     href: The raw href from the markdown link, if any.
- *     artifacts: Artifacts this turn produced, in the order the tool made them
- *         (``A1`` is the first).
+ *     artifacts: Every artifact in the conversation, in creation order (``A1``
+ *         is the first). Refs are conversation-scoped, so a link may point at
+ *         a file an earlier turn produced.
+ *     turnArtifacts: The artifacts of the turn being rendered, when known.
+ *         Only the filename fallback consults it; see below.
  *
  * Returns:
  *     How the link should be rendered. ``external`` for ordinary URLs,
@@ -133,6 +136,7 @@ function decodeSegment(segment: string): string {
 export function resolveSandboxLink(
   href: string | undefined | null,
   artifacts: SandboxArtifact[] | undefined,
+  turnArtifacts?: SandboxArtifact[],
 ): SandboxLinkResolution {
   if (typeof href !== 'string' || !href) return { kind: 'external' };
   const scheme = matchedScheme(href);
@@ -177,11 +181,18 @@ export function resolveSandboxLink(
   // Fabricated paths (``/mnt/data/…``, ``/tmp/…``, a bare filename) — and
   // ``/artifact/<filename>``, which the model also writes — usually still name
   // a file that exists as an artifact on this turn.
-  const byName = findLast(
-    available,
-    (artifact) =>
-      !!artifact.label && artifact.label.toLowerCase() === last.toLowerCase(),
-  );
+  // This turn's own files win. An id or a ref names exactly one artifact, so
+  // those match conversation-wide above; a filename does not — two turns can
+  // each produce ``report.pdf``, and `_filename()` hands EVERY html artifact
+  // that same name. Resolving an old message's link to the newest copy made
+  // the prose link disagree with the download chip in its own bubble, which
+  // renders from that turn's artifacts alone. Falls back to the conversation
+  // so a link naming a file from an earlier turn still resolves.
+  const nameMatches = (artifact: SandboxArtifact) =>
+    !!artifact.label && artifact.label.toLowerCase() === last.toLowerCase();
+  const byName =
+    findLast(turnArtifacts ?? [], nameMatches) ??
+    findLast(available, nameMatches);
   if (byName) return { kind: 'artifact', artifact: byName };
 
   return miss;
diff --git a/tests/api/answer/routes/test_answer.py b/tests/api/answer/routes/test_answer.py
index 00d93b6b..4b9393b4 100644
--- a/tests/api/answer/routes/test_answer.py
+++ b/tests/api/answer/routes/test_answer.py
@@ -143,6 +143,47 @@ class TestAnswerResourcePost:
             assert resp.status_code == 400
             assert resp.get_json()["error"] == "Stream error"
 
+    def test_resume_already_in_progress_returns_409(
+        self, answer_client, mock_stream_processor,
+    ):
+        """A raced resume is a conflict the caller can retry, not a 500.
+
+        ``ResumeInProgressError`` subclasses ``ValueError`` but reaches this
+        route through ``resume_from_tool_actions``, where the bare
+        ``except Exception`` reported it as an opaque 500 with a full
+        traceback on the ERROR channel. ``/stream`` and
+        ``/v1/chat/completions`` both answer 409 for the identical condition.
+        """
+        from application.api.answer.services.continuation_service import (
+            ResumeInProgressError,
+        )
+
+        mock_stream_processor.resume_from_tool_actions.side_effect = (
+            ResumeInProgressError("Resume already in progress for this conversation.")
+        )
+
+        with patch(
+            "application.api.answer.routes.answer.AnswerResource.validate_request",
+            return_value=None,
+        ):
+            resp = answer_client.post(
+                "/api/answer",
+                data=json.dumps({
+                    "question": "",
+                    "conversation_id": _CONV_ID,
+                    "tool_actions": [{"call_id": "call_1", "decision": "approved"}],
+                }),
+                content_type="application/json",
+            )
+
+        assert resp.status_code == 409
+        body = resp.get_json()
+        assert body["code"] == "resume_in_progress"
+        assert "Resume already in progress" in body["error"]
+        # The generic handler's opaque text would tell the caller nothing
+        # about retryability.
+        assert body["error"] != "An error occurred processing your request"
+
     def test_exception_returns_500(self, answer_client, mock_stream_processor):
         with patch(
             "application.api.answer.routes.answer.AnswerResource.validate_request",
diff --git a/tests/api/answer/test_base_routes.py b/tests/api/answer/test_base_routes.py
index dac6b2c4..abe14983 100644
--- a/tests/api/answer/test_base_routes.py
+++ b/tests/api/answer/test_base_routes.py
@@ -1450,3 +1450,67 @@ class TestStreamingHeartbeatSeed:
             )
 
             resource.conversation_service.heartbeat_message.assert_not_called()
+
+
+class TestCompleteStreamSupersededLookupIsGuarded:
+    """A DB failure in the supersede lookup must not error a finished turn.
+
+    The ``NOT_FOUND`` branch used to only log. It now consults
+    ``was_superseded`` to tell a replaced turn from an orphaned answer — over
+    its own connection, inside the streaming ``try``. Unguarded, a blip there
+    (or a deploy where the app runs ahead of migration 0030 and the
+    ``superseded_messages`` table does not exist yet) escapes to the generic
+    handler: the client gets a terminal ``error`` after the whole answer, the
+    row is finalized a second time, a continuation's resume claim is released,
+    and the ``answer_persist_failed`` alert this branch exists to raise is
+    swallowed.
+    """
+
+    def test_lookup_failure_falls_back_instead_of_erroring_the_stream(
+        self, pg_conn, flask_app, caplog,
+    ):
+        import logging
+
+        from application.api.answer.routes.base import BaseAnswerResource
+        from application.api.answer.services.conversation_service import (
+            ConversationService,
+        )
+        from application.storage.db.repositories.conversations import (
+            MessageUpdateOutcome,
+        )
+
+        with flask_app.app_context():
+            resource = BaseAnswerResource()
+            mock_agent = MagicMock()
+            mock_agent.gen.return_value = iter(
+                [{"type": "answer", "answer": "the answer"}]
+            )
+
+            with _patch_db_session(pg_conn), patch.object(
+                ConversationService,
+                "finalize_message",
+                return_value=MessageUpdateOutcome.NOT_FOUND,
+            ), patch.object(
+                ConversationService,
+                "was_superseded",
+                side_effect=RuntimeError("server closed the connection"),
+            ), caplog.at_level(logging.ERROR):
+                stream = list(
+                    resource.complete_stream(
+                        question="does the guard hold?",
+                        agent=mock_agent,
+                        conversation_id=None,
+                        user_api_key=None,
+                        decoded_token={"sub": "u-superseded-guard"},
+                        should_persist=True,
+                        model_id="gpt-4",
+                    )
+                )
+
+        # The answer was already streamed; the row is gone either way. What
+        # must not happen is retracting it with a terminal error frame.
+        assert [s for s in stream if '"type": "error"' in s] == []
+        # And the alert must still fire, on the orphaned-answer path.
+        assert any(
+            "answer_persist_failed" in r.getMessage() for r in caplog.records
+        )
diff --git a/tests/api/user/test_reconciliation.py b/tests/api/user/test_reconciliation.py
index db97bda5..e59df5a2 100644
--- a/tests/api/user/test_reconciliation.py
+++ b/tests/api/user/test_reconciliation.py
@@ -945,17 +945,20 @@ class TestStalledIngestEvent:
 class TestPostgresUriMissing:
     @pytest.mark.unit
     def test_returns_skip_dict(self, monkeypatch):
-        from application.api.user.reconciliation import run_reconciliation
+        from application.api.user.reconciliation import (
+            run_reconciliation,
+            zero_summary,
+        )
         from application.core.settings import settings
 
         monkeypatch.setattr(settings, "POSTGRES_URI", None, raising=False)
 
         result = run_reconciliation()
-        assert result == {
-            "messages_failed": 0,
-            "tool_calls_failed": 0,
-            "skipped": "POSTGRES_URI not set",
-        }
+        # Same counters as any other tick, all zero, plus the reason. Pinning
+        # a hand-written literal here is what let the shapes drift apart: this
+        # path reported 3 of the 5 keys and the task's error fallback invented
+        # a sixth that no sweep produces.
+        assert result == {**zero_summary(), "skipped": "POSTGRES_URI not set"}
 
 
 # ---------------------------------------------------------------------------
diff --git a/tests/api/user/test_tasks.py b/tests/api/user/test_tasks.py
index 59982227..1624a71a 100644
--- a/tests/api/user/test_tasks.py
+++ b/tests/api/user/test_tasks.py
@@ -900,3 +900,38 @@ class TestParseTimeoutForSize:
         assert limits["time_limit"] - limits["soft_time_limit"] == grace
         # Never zero/negative, whatever the caller computed.
         assert parse_task_time_limits(0)["soft_time_limit"] == 1
+
+
+class TestReconciliationTaskShape:
+    """The beat's error fallback must report the same counters as a good tick."""
+
+    @pytest.mark.unit
+    def test_error_fallback_matches_the_real_summary_keys(self):
+        from application.api.user import reconciliation
+        from application.api.user.tasks import reconciliation_task
+
+        with patch.object(
+            reconciliation, "run_reconciliation", side_effect=RuntimeError("db down")
+        ):
+            result = reconciliation_task.run()
+
+        assert result["error"] is True
+        # Hand-writing the fallback drifted immediately: it invented
+        # ``attachments_stalled`` (no sweep produces it) and dropped
+        # ``idempotency_pending_failed``, so a failed tick reported counters
+        # that could never appear and hid one that can.
+        counters = {k: v for k, v in result.items() if k != "error"}
+        assert counters == reconciliation.zero_summary()
+        assert all(v == 0 for v in counters.values())
+
+    @pytest.mark.unit
+    def test_skipped_tick_reports_the_same_counters(self, monkeypatch):
+        from application.api.user import reconciliation
+
+        monkeypatch.setattr(reconciliation.settings, "POSTGRES_URI", "", raising=False)
+        result = reconciliation.run_reconciliation()
+
+        assert result["skipped"] == "POSTGRES_URI not set"
+        assert {k: v for k, v in result.items() if k != "skipped"} == (
+            reconciliation.zero_summary()
+        )
diff --git a/tests/core/test_logging_config.py b/tests/core/test_logging_config.py
index cff033c5..63ea0512 100644
--- a/tests/core/test_logging_config.py
+++ b/tests/core/test_logging_config.py
@@ -102,3 +102,65 @@ class TestSetupLogging:
 
         handlers = logging.getLogger().handlers
         assert otel_handler not in handlers
+
+
+class TestAlembicDoesNotSilenceApplicationLoggers:
+    """A boot that applies a migration must not switch off app logging.
+
+    ``alembic/env.py`` calls ``logging.config.fileConfig``, whose
+    ``disable_existing_loggers`` defaults to True and whose ini names only
+    ``root``/``sqlalchemy``/``alembic`` — so every ``application.*`` logger
+    imported before it is disabled. ``app.py`` calls ``setup_logging()`` at
+    line 16 but ``ensure_database_ready()`` at line 58, so the web tier never
+    re-enables them: on the one boot where a schema upgrade happened, it
+    logs nothing for the rest of the process's life. k8s sets
+    ``AUTO_MIGRATE=false``, but every docker-compose variant leaves it on.
+    """
+
+    @pytest.mark.unit
+    def test_env_py_opts_out_of_disable_existing_loggers(self):
+        """AST, not a grep: the kwarg has to be on the real call."""
+        import ast
+        import pathlib
+
+        source = pathlib.Path("application/alembic/env.py").read_text()
+        calls = [
+            node
+            for node in ast.walk(ast.parse(source))
+            if isinstance(node, ast.Call)
+            and getattr(node.func, "id", None) == "fileConfig"
+        ]
+        assert calls, "env.py no longer calls fileConfig"
+        for call in calls:
+            kwargs = {kw.arg: kw.value for kw in call.keywords}
+            assert "disable_existing_loggers" in kwargs
+            assert kwargs["disable_existing_loggers"].value is False
+
+    @pytest.mark.unit
+    def test_the_flag_is_what_keeps_application_loggers_alive(self, tmp_path):
+        """Pins the mechanism, in a subprocess so pytest's own logging survives.
+
+        ``fileConfig`` mutates global logging state, so this cannot run
+        in-process without wrecking every later test's log capture.
+        """
+        import subprocess
+        import sys
+
+        probe = tmp_path / "probe.py"
+        probe.write_text(
+            "import logging\n"
+            "from logging.config import fileConfig\n"
+            "log = logging.getLogger('application.api.answer.routes.stream')\n"
+            "fileConfig('application/alembic.ini', disable_existing_loggers=True)\n"
+            "kept = logging.getLogger('application.api.answer.routes.stream')\n"
+            "print('default_disables', kept.disabled)\n"
+            "kept.disabled = False\n"
+            "fileConfig('application/alembic.ini', disable_existing_loggers=False)\n"
+            "print('flag_preserves', not kept.disabled)\n"
+        )
+        out = subprocess.run(
+            [sys.executable, str(probe)],
+            capture_output=True, text=True, check=True,
+        ).stdout
+        assert "default_disables True" in out
+        assert "flag_preserves True" in out
diff --git a/tests/e2e/specs/tier-a/resume-retry-persists.spec.ts b/tests/e2e/specs/tier-a/resume-retry-persists.spec.ts
new file mode 100644
index 00000000..8744b47e
--- /dev/null
+++ b/tests/e2e/specs/tier-a/resume-retry-persists.spec.ts
@@ -0,0 +1,259 @@
+/**
+ * Retrying a failed resume must keep the answer it produces.
+ *
+ * `release_claim` hands a failed resume's `pending_tool_state` row back to
+ * `pending` so the user can retry immediately instead of waiting out
+ * `revert_stale_resuming`'s 600 s grace. But the same `except Exception` in
+ * `routes/base.py` also finalizes the reserved message row `failed`, and a
+ * resume reuses the SAME `reserved_message_id` — `resume_from_tool_actions`
+ * reads it back out of the persisted `agent_config`
+ * (services/stream_processor.py). So the retry streams into a row that is
+ * already terminal, and `finalize_message(status="complete")` is gated by
+ * `only_if_non_terminal`, whose reclaim hole originally matched only the
+ * reconciler's own error string. Outcome: `ALREADY_FAILED`.
+ *
+ * // Silent-break covered: the user watches a correct answer stream in, then
+ *    on reload finds "Response was terminated prior to completion" with no way
+ *    back to it. Nothing in the SSE body reveals the loss — the wire shows a
+ *    perfectly healthy turn — so these specs assert on the DB row and on
+ *    GET /api/get_single_conversation, never on the stream.
+ *
+ * The fix stamps `resume_retryable` on the failure exactly when the claim was
+ * released, and widens the reclaim gate to honour it. The two side effects
+ * have to agree: releasing the claim says "retryable", so the row must not
+ * simultaneously be terminal in a way the retry cannot overwrite.
+ */
+
+import * as playwright from '@playwright/test';
+const { expect, test } = playwright;
+
+import { authedRequest } from '../../helpers/api.js';
+import { pg } from '../../helpers/db.js';
+import { resetDb } from '../../helpers/reset.js';
+import { newUserContext } from '../../helpers/auth.js';
+import { streamOnce } from '../../helpers/streaming.js';
+
+/** application/api/answer/services/conversation_service.py */
+const TERMINATED_PLACEHOLDER =
+  'Response was terminated prior to completion, try regenerating.';
+
+interface MessageRow {
+  id: string;
+  status: string;
+  response: string;
+  message_metadata: Record<string, unknown> | null;
+}
+
+async function messageRow(id: string): Promise<MessageRow> {
+  const { rows } = await pg.query<MessageRow>(
+    `SELECT id, status, response, message_metadata
+       FROM conversation_messages WHERE id = CAST($1 AS uuid)`,
+    [id],
+  );
+  return rows[0];
+}
+
+async function claimStatus(convId: string): Promise<string | null> {
+  const { rows } = await pg.query<{ status: string }>(
+    `SELECT status FROM pending_tool_state
+      WHERE conversation_id = CAST($1 AS uuid)`,
+    [convId],
+  );
+  return rows.length ? rows[0].status : null;
+}
+
+/**
+ * Park a message row mid-flight and point a claimable `pending_tool_state` at
+ * it, exactly as a turn paused on a client-side tool does. `agent_config`
+ * carries the `reserved_message_id`, which is what makes the retry reuse this
+ * same row.
+ */
+async function seedPausedTurn(
+  convId: string,
+  sub: string,
+): Promise<string> {
+  const { rows } = await pg.query<{ id: string }>(
+    `INSERT INTO conversation_messages (
+        conversation_id, position, prompt, response, status, user_id,
+        timestamp, message_metadata
+     ) VALUES (CAST($1 AS uuid), 99, 'paused on a tool', '', 'streaming', $2,
+               clock_timestamp(), '{}'::jsonb)
+     RETURNING id`,
+    [convId, sub],
+  );
+  const messageId = rows[0].id;
+
+  await pg.query(
+    `INSERT INTO pending_tool_state (
+        conversation_id, user_id, messages, pending_tool_calls,
+        tools_dict, tool_schemas, agent_config, expires_at, status
+     ) VALUES (
+        CAST($1 AS uuid), $2,
+        CAST($3 AS jsonb), CAST($4 AS jsonb),
+        '{}'::jsonb, '[]'::jsonb, CAST($5 AS jsonb),
+        clock_timestamp() + interval '30 minutes', 'pending')`,
+    [
+      convId,
+      sub,
+      JSON.stringify([
+        { role: 'user', content: 'paused on a tool' },
+        {
+          role: 'assistant',
+          content: null,
+          // No `name` on the function — enough to make gen_continuation raise,
+          // which is the "resume failed once" trigger this spec is about.
+          tool_calls: [
+            { id: 'call_1', type: 'function', function: { arguments: '{}' } },
+          ],
+        },
+      ]),
+      JSON.stringify([
+        {
+          call_id: 'call_1',
+          tool_name: 'noop',
+          action_name: 'noop',
+          arguments: {},
+        },
+      ]),
+      JSON.stringify({
+        llm_name: 'openai',
+        agent_type: 'ClassicAgent',
+        prompt: 'You are a test assistant.',
+        reserved_message_id: messageId,
+      }),
+    ],
+  );
+  return messageId;
+}
+
+test.describe('retrying a failed resume', () => {
+  test.beforeEach(async () => {
+    await resetDb();
+  });
+
+  test('a failed resume hands the claim back instead of holding it for 600s', async ({
+    browser,
+  }) => {
+    const { sub, token } = await newUserContext(browser);
+    const api = await authedRequest(playwright, token);
+    try {
+      const convId = await streamOnce(api, {
+        question: 'hello',
+        save_conversation: true,
+      });
+      await seedPausedTurn(convId, sub);
+
+      const res = await api.post('/stream', {
+        data: {
+          question: '',
+          conversation_id: convId,
+          tool_actions: [{ call_id: 'call_1', decision: 'approved' }],
+        },
+      });
+      await res.text();
+
+      // Not `resuming`: that is the ten-minute lockout from the 2026-08-19
+      // incident, recoverable only by the janitor.
+      expect(await claimStatus(convId)).not.toBe('resuming');
+    } finally {
+      await api.dispose();
+    }
+  });
+
+  test('the failure it leaves behind is marked retryable, not plainly terminal', async ({
+    browser,
+  }) => {
+    const { sub, token } = await newUserContext(browser);
+    const api = await authedRequest(playwright, token);
+    try {
+      const convId = await streamOnce(api, {
+        question: 'hello',
+        save_conversation: true,
+      });
+      const messageId = await seedPausedTurn(convId, sub);
+
+      const res = await api.post('/stream', {
+        data: {
+          question: '',
+          conversation_id: convId,
+          tool_actions: [{ call_id: 'call_1', decision: 'approved' }],
+        },
+      });
+      await res.text();
+
+      const row = await messageRow(messageId);
+      expect(row.status).toBe('failed');
+      expect(row.response).toBe(TERMINATED_PLACEHOLDER);
+      // The marker the reclaim gate keys on. Without it the retry below can
+      // never overwrite this row, and its answer is discarded in silence.
+      expect(row.message_metadata?.resume_retryable).toBe(true);
+    } finally {
+      await api.dispose();
+    }
+  });
+
+  test("a retry's answer replaces the failed row instead of being discarded", async ({
+    browser,
+  }) => {
+    const { sub, token } = await newUserContext(browser);
+    const api = await authedRequest(playwright, token);
+    try {
+      const convId = await streamOnce(api, {
+        question: 'hello',
+        save_conversation: true,
+      });
+      const messageId = await seedPausedTurn(convId, sub);
+
+      // Resume #1 — errors, releases the claim, marks the row retryable.
+      const first = await api.post('/stream', {
+        data: {
+          question: '',
+          conversation_id: convId,
+          tool_actions: [{ call_id: 'call_1', decision: 'approved' }],
+        },
+      });
+      await first.text();
+      expect((await messageRow(messageId)).status).toBe('failed');
+
+      // Resume #2 — the row is claimable again, so re-approving works. Repair
+      // the stored messages first: the retry is the user clicking approve a
+      // second time on a turn whose state is intact, not a replay of the
+      // corruption that broke resume #1.
+      await pg.query(
+        `UPDATE pending_tool_state
+            SET messages = jsonb_set(
+                  messages, '{1,tool_calls,0,function,name}', '"noop"'::jsonb)
+          WHERE conversation_id = CAST($1 AS uuid)`,
+        [convId],
+      );
+
+      const second = await api.post('/stream', {
+        data: {
+          question: '',
+          conversation_id: convId,
+          tool_actions: [{ call_id: 'call_1', decision: 'approved' }],
+        },
+      });
+      const body = await second.text();
+      expect(second.status()).toBe(200);
+
+      // The whole point: whatever the second attempt produced must be what the
+      // row holds. A silently-dropped answer leaves the placeholder behind.
+      const row = await messageRow(messageId);
+      if (body.includes('"type": "end"')) {
+        expect(row.status).toBe('complete');
+        expect(row.response).not.toBe(TERMINATED_PLACEHOLDER);
+        // Reclaim strips the failure bookkeeping, or a finished answer
+        // carries a stale error into the API response.
+        expect(row.message_metadata?.error).toBeUndefined();
+        expect(row.message_metadata?.resume_retryable).toBeUndefined();
+      } else {
+        // If the second attempt also failed, it must at least still be
+        // retryable — never wedged terminal with the claim gone.
+        expect(row.message_metadata?.resume_retryable).toBe(true);
+      }
+    } finally {
+      await api.dispose();
+    }
+  });
+});
diff --git a/tests/e2e/specs/tier-b/streaming-ux.spec.ts b/tests/e2e/specs/tier-b/streaming-ux.spec.ts
index 2c970d27..3e5f5bbf 100644
--- a/tests/e2e/specs/tier-b/streaming-ux.spec.ts
+++ b/tests/e2e/specs/tier-b/streaming-ux.spec.ts
@@ -174,13 +174,21 @@ test.describe('tier-b · streaming UX', () => {
     // We model the tool-approval flow at the table level — the contract is
     // that the route path with `tool_actions` + `conversation_id` calls
     // `StreamProcessor.resume_from_tool_actions`, which loads from PG, then
-    // `mark_resuming` flips status `pending`→`resuming` so the row cannot
-    // be re-claimed (see continuation_service.py + repositories/pending_tool_state.py).
-    // On clean completion the route layer DELETEs the row
-    // (routes/base.py: `delete_state` after stream success); on a crash the
-    // row stays in `resuming` for the janitor to revert after the grace
-    // window. Either outcome proves the replay footgun is closed — what
-    // must not happen is the row remaining in `pending` status.
+    // claims the row so it cannot be claimed twice
+    // (see continuation_service.py + repositories/pending_tool_state.py).
+    //
+    // What the claim must never be is STUCK. It originally could only be
+    // released by `revert_stale_resuming`'s 600 s grace, which locked the user
+    // out of their own conversation for ten minutes after any resume that
+    // failed. `release_claim` now hands the row back on the error path, so the
+    // outcomes are: DELETED (clean completion, `delete_state`), `pending`
+    // (released for an immediate retry), or `resuming` only while a resume is
+    // genuinely still in flight.
+    //
+    // A released row IS re-fireable, which is the point — the replay footgun
+    // that justified the old `pending`-is-forbidden rule is now handled by the
+    // owning message row instead: see `resume-retry-persists.spec.ts`, which
+    // pins that the retry lands on the same row and its answer survives.
     // No real agent wiring is needed to test that contract.
     const { sub, token } = await newUserContext(browser);
     const api = await authedRequest(playwright, token);
@@ -278,15 +286,15 @@ test.describe('tier-b · streaming UX', () => {
       expect(resume.status()).toBe(200);
       await resume.text();
 
-      // 4. Row must be either deleted (clean completion) or marked
-      // `resuming` (crash partway). What must not happen is the row
-      // remaining in `pending` — that would mean the resume is
-      // re-fireable, a footgun mirroring Mongo-era bug #2088.
-      const pendingCount = await countRows('pending_tool_state', {
-        sql: "user_id = $1 AND conversation_id = CAST($2 AS uuid) AND status = 'pending'",
+      // 4. The resume errored (the mock LLM cannot continue from a fabricated
+      // tool_call), so the claim must have been handed back rather than left
+      // held. Either the row is gone or it is `pending` and immediately
+      // retryable; a leftover `resuming` is the ten-minute lockout.
+      const stuckCount = await countRows('pending_tool_state', {
+        sql: "user_id = $1 AND conversation_id = CAST($2 AS uuid) AND status = 'resuming'",
         params: [sub, convId],
       });
-      expect(pendingCount).toBe(0);
+      expect(stuckCount).toBe(0);
     } finally {
       await api.dispose();
     }
diff --git a/tests/storage/db/repositories/test_reconciliation_liveness.py b/tests/storage/db/repositories/test_reconciliation_liveness.py
index cf579e05..e8a04e48 100644
--- a/tests/storage/db/repositories/test_reconciliation_liveness.py
+++ b/tests/storage/db/repositories/test_reconciliation_liveness.py
@@ -278,6 +278,56 @@ class TestFinalizeReclaim:
         assert "error" not in meta
         assert "reconcile_attempts" not in meta
 
+    def test_reclaims_a_resume_that_released_its_claim(self, pg_conn):
+        """A failed resume that handed the turn back must not block the retry.
+
+        The stream handler releases the continuation claim and stamps
+        ``resume_retryable`` in the same breath. The retry reuses this very row
+        (the reserved id lives in the persisted ``agent_config``), so without
+        the second reclaim hole its successful finalize lands ALREADY_FAILED
+        and the user's answer is discarded after they watched it stream in.
+        """
+        msg = _seed_message(
+            pg_conn,
+            status="failed",
+            metadata={"error": "KeyError: 'name'", "resume_retryable": True},
+        )
+        repo = ConversationsRepository(pg_conn)
+
+        outcome = repo.update_message_by_id(
+            msg["id"],
+            {"response": "the retry's answer", "status": "complete"},
+            only_if_non_terminal=True,
+            reclaim_reconciler_failed=True,
+        )
+
+        assert outcome is MessageUpdateOutcome.UPDATED
+        assert _status(pg_conn, msg["id"]) == "complete"
+        # The marker must not survive onto the finished row, or the API
+        # response carries a stale failure alongside a complete answer.
+        meta = _meta(pg_conn, msg["id"])
+        assert "resume_retryable" not in meta
+        assert "error" not in meta
+
+    def test_does_not_reclaim_a_failure_that_kept_its_claim(self, pg_conn):
+        """No marker means no release happened, so the row stays terminal."""
+        msg = _seed_message(
+            pg_conn,
+            status="failed",
+            metadata={"error": "KeyError: 'name'"},
+        )
+        repo = ConversationsRepository(pg_conn)
+
+        outcome = repo.update_message_by_id(
+            msg["id"],
+            {"response": "late answer", "status": "complete"},
+            only_if_non_terminal=True,
+            reclaim_reconciler_failed=True,
+        )
+
+        assert outcome is MessageUpdateOutcome.ALREADY_FAILED
+        assert _status(pg_conn, msg["id"]) == "failed"
+
     def test_does_not_reclaim_genuine_failure(self, pg_conn):
         """Client disconnect, provider errors etc. stay terminal."""
         msg = _seed_message(
diff --git a/tests/test_celery.py b/tests/test_celery.py
index f6d3477b..8994dbc4 100644
--- a/tests/test_celery.py
+++ b/tests/test_celery.py
@@ -128,3 +128,97 @@ def test_durable_task_retry_envelope_spans_a_multi_minute_outage():
     assert 150 <= statistics.median(envelopes) <= 270
     # A 60 s outage must be survivable in the large majority of retry runs.
     assert sum(e >= 60 for e in envelopes) / len(envelopes) > 0.9
+
+
+@pytest.mark.unit
+def test_durable_tasks_never_retry_a_deterministic_parse_failure():
+    """The widened envelope needs ``dont_autoretry_for`` as its counterweight.
+
+    ``autoretry_for=(Exception,)`` retries EVERYTHING, so raising the backoff
+    factor to 60 also stretched permanent failures — an empty, image-only or
+    unparseable file — from a ~7 s envelope to a median 3.5 min and up to ~7.
+    Anything polling ``/api/task_status`` (the wiki-convert and GraphRAG-enable
+    modals) reports "pending" for that whole time, and the work re-runs four
+    times to fail identically. ``e2e specs/tier-b/upload.spec.ts`` went red on
+    exactly this.
+    """
+    from unittest.mock import MagicMock
+
+    from celery.app.autoretry import add_autoretry_behaviour
+
+    from application.api.user.tasks import DURABLE_TASK
+    from application.parser.file.base_parser import DocumentParseError
+
+    assert DocumentParseError in DURABLE_TASK["dont_autoretry_for"]
+
+    retried: list[dict] = []
+
+    class _Task:
+        max_retries = DURABLE_TASK["max_retries"]
+
+        def __init__(self) -> None:
+            self.request = MagicMock()
+            self.request.retries = 0
+
+        def run(self):
+            raise DocumentParseError("No text could be extracted from this file.")
+
+        def retry(self, **kwargs):
+            retried.append(dict(kwargs))
+            return RuntimeError("retry-sentinel")
+
+    task = _Task()
+    add_autoretry_behaviour(
+        task,
+        autoretry_for=DURABLE_TASK["autoretry_for"],
+        dont_autoretry_for=DURABLE_TASK["dont_autoretry_for"],
+        retry_kwargs={"max_retries": DURABLE_TASK["max_retries"]},
+        retry_backoff=DURABLE_TASK["retry_backoff"],
+    )
+
+    # Propagates unchanged to the poison/failure path instead of being
+    # swallowed into a retry.
+    with pytest.raises(DocumentParseError):
+        task.run()
+    assert retried == []
+
+
+@pytest.mark.unit
+def test_every_durable_task_carries_the_parse_failure_guard():
+    """Registered tasks, not just the shared dict — a decorator can override it.
+
+    ``store_attachment`` passes a wider tuple through ``durable_task()``; the
+    point is that no durable task ends up with a NARROWER one, which is how
+    seven of the nine came to retry a permanent parse failure four times.
+    """
+    from application.api.user import tasks as user_tasks
+    from application.parser.file.base_parser import DocumentParseError
+
+    durable = (
+        "ingest",
+        "ingest_remote",
+        "reingest_source_task",
+        "reembed_wiki_page",
+        "convert_source_to_wiki",
+        "extract_graph",
+        "process_agent_webhook",
+        "ingest_connector_task",
+        "store_attachment",
+    )
+    for name in durable:
+        task = getattr(user_tasks, name)
+        assert DocumentParseError in getattr(task, "dont_autoretry_for", ()), name
+
+
+@pytest.mark.unit
+def test_unparseable_file_raises_the_non_retryable_type():
+    """The guard is only reachable if the pipeline raises the right class.
+
+    This used to be a bare ``ValueError``, which ``autoretry_for=(Exception,)``
+    swept up regardless of the ``dont_autoretry_for`` tuple.
+    """
+    from application.parser.embedding_pipeline import embed_and_store_documents
+    from application.parser.file.base_parser import DocumentParseError
+
+    with pytest.raises(DocumentParseError, match="No text could be extracted"):
+        embed_and_store_documents([], "/tmp", "src", None)
diff --git a/tests/test_coverage_gaps.py b/tests/test_coverage_gaps.py
index d3aa3028..e2ebfbb7 100644
--- a/tests/test_coverage_gaps.py
+++ b/tests/test_coverage_gaps.py
@@ -845,8 +845,9 @@ class TestEmbeddingPipeline:
 
     def test_embed_and_store_empty_docs(self):
         from application.parser.embedding_pipeline import embed_and_store_documents
+        from application.parser.file.base_parser import DocumentParseError
 
-        with pytest.raises(ValueError, match="No text could be extracted"):
+        with pytest.raises(DocumentParseError, match="No text could be extracted"):
             embed_and_store_documents([], "/tmp/test", "src1", MagicMock())
 
     def test_embed_and_store_creates_folder(self, tmp_path):
diff --git a/tests/test_remaining_coverage.py b/tests/test_remaining_coverage.py
index ff5befe6..3bbaad44 100644
--- a/tests/test_remaining_coverage.py
+++ b/tests/test_remaining_coverage.py
@@ -862,10 +862,11 @@ class TestEmbeddingPipelineGaps:
                 assert os.path.exists(folder)
 
     def test_embed_and_store_raises_on_empty_docs(self):
-        """Cover line 69: raises ValueError when docs is empty."""
+        """Cover line 69: raises DocumentParseError when docs is empty."""
         from application.parser.embedding_pipeline import embed_and_store_documents
+        from application.parser.file.base_parser import DocumentParseError
 
-        with pytest.raises(ValueError, match="No text could be extracted"):
+        with pytest.raises(DocumentParseError, match="No text could be extracted"):
             embed_and_store_documents([], "/tmp/test", "source_id", MagicMock())
 
 
@@ -1042,10 +1043,11 @@ class TestEmbeddingPipelineCoverage:
         assert sanitize_content(None) is None
 
     def test_embed_and_store_empty_docs_raises(self, tmp_path):
-        """Cover line 69: empty docs raises ValueError."""
+        """Cover line 69: empty docs raises DocumentParseError."""
         from application.parser.embedding_pipeline import embed_and_store_documents
+        from application.parser.file.base_parser import DocumentParseError
 
-        with pytest.raises(ValueError, match="No text could be extracted"):
+        with pytest.raises(DocumentParseError, match="No text could be extracted"):
             embed_and_store_documents([], str(tmp_path / "test"), "src-1", None)
 
     def test_embed_and_store_creates_folder(self, tmp_path):
@@ -1241,10 +1243,11 @@ class TestBlankDocumentsAreRejected:
     @pytest.mark.parametrize("blank", ["", "   ", "\n\t  \n"])
     def test_whitespace_only_document_is_rejected(self, blank, tmp_path):
         from application.parser.embedding_pipeline import embed_and_store_documents
+        from application.parser.file.base_parser import DocumentParseError
 
         class _Doc:
             def __init__(self, text):
                 self.text = text
 
-        with pytest.raises(ValueError, match="No text could be extracted"):
+        with pytest.raises(DocumentParseError, match="No text could be extracted"):
             embed_and_store_documents([_Doc(blank)], str(tmp_path), "src-1", None)
diff --git a/tests/vectorstore/test_pgvector_schema_and_pool.py b/tests/vectorstore/test_pgvector_schema_and_pool.py
index ba4d499c..072dd270 100644
--- a/tests/vectorstore/test_pgvector_schema_and_pool.py
+++ b/tests/vectorstore/test_pgvector_schema_and_pool.py
@@ -9,6 +9,8 @@ one-shot safety net for processes that never ran the boot hook.
 
 from __future__ import annotations
 
+import gc
+
 from unittest.mock import MagicMock, Mock, patch
 
 import pytest
@@ -266,6 +268,16 @@ class TestReadPathsCloseTheirTransaction:
 @pytest.mark.unit
 class TestConnectionPooling:
     def _pooled_store(self, monkeypatch):
+        # ``PGVectorStore.__del__`` releases into the module-global
+        # ``_POOLS``, and a store left holding a connection by an earlier test
+        # is only collectable through a MagicMock reference cycle — so its
+        # finalizer fires at an arbitrary later moment. If that moment lands
+        # inside a test that has patched ``_POOLS[dsn]`` to its own fake pool,
+        # the stale release is recorded against THIS test's mock and
+        # ``putconn.assert_called_once_with`` sees two calls. Drain pending
+        # finalizers here, before any patching, so the count is deterministic
+        # whatever ran before.
+        gc.collect()
         store, _, _, _ = _make_store()
         store._connection = None
         store._pool_max_size = 4