fix(downloader): don't misclassify tier-gated runs with per-item video 403s

When a Patreon source loses tier access, gallery-dl emits "Not allowed to
view post N" warnings for every post AND yt-dlp tries to fetch HLS manifests
which also 403. The ytdl error text contained "HTTP Error 403: Forbidden",
which the classifier's ACCESS_DENIED_PATTERNS matched against the full
combined stdout+stderr — so the run was labeled ACCESS_DENIED instead of
TIER_LIMITED.

Two fixes in one edit, because they're interlocked:
- Compute per-item vs source-level error lines upfront. has_actual_error
  now reflects only source-level errors (previously the per-item exclusion
  was gated on skip evidence, which tier-limited runs don't produce since
  no content was accessed).
- Strip per-item error lines from `combined` before downstream pattern
  matching so noise from recovered per-item failures doesn't latch onto
  ACCESS_DENIED / HTTP_ERROR / NOT_FOUND classifiers.

Regression test: tier-gated Patreon run with yt-dlp HLS 403s → TIER_LIMITED,
not ACCESS_DENIED.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

backend/app/services/gallery_dl.py

@@ -437,40 +437,52 @@ class GalleryDLService:
             "download failed",
             "extraction failed",
         ]
-        has_actual_error = any(err in combined for err in actual_error_indicators)
 
-        # Non-fatal error lines gallery-dl logs but recovers from. These shouldn't
+        # Non-fatal error lines gallery-dl logs but recovers from. These are
-        # block skip detection when the overall download proceeded successfully:
+        # per-item failures the run already moved past — they shouldn't cause
-        #   - "not allowed to view" / "unable to get post": per-item paywall warnings
+        # source-level misclassification:
-        #     on Patreon /c/user/posts URLs where some posts are paywalled.
+        #   - "not allowed to view" / "unable to get post": per-item paywall
-        #   - "failed to extract campaign id": logged when the vanity-URL HTML path
+        #     warnings on Patreon URLs with mixed tier access.
-        #     fails, but gallery-dl falls back to the /cw/<vanity> redirect and
+        #   - "failed to extract campaign id": vanity-URL HTML path fails,
-        #     recovers the campaign_id on its own.
+        #     gallery-dl falls back to /cw/<vanity> and recovers.
-        #   - "failed to download": per-item download failure (e.g., a single
+        #   - "failed to download": per-item media failure (expired URL,
-        #     Patreon attachment whose media URL expired / was deleted). The
+        #     yt-dlp HLS 403 on tier-gated video, deleted attachment).
-        #     downloader logs [download][error] and moves to the next item.
+        #     Downloader logs [download][error] and continues.
-        if has_actual_error and (skip_line_count > 0 or has_skip_text):
+        #   - "cannot import yt-dlp"/"youtube-dl": video backend missing;
+        #     surfaces via run_stats as a setup gap, not a source failure.
         per_item_patterns = [
             "not allowed to view",
             "unable to get post",
             "failed to extract campaign id",
             "failed to download",
-                # yt-dlp/youtube-dl missing: each video that gallery-dl tries to
-                # download fails, but image/skip activity continues. Treat as
-                # per-item so it doesn't mask TIER_LIMITED/NO_NEW_CONTENT.
-                # Surface separately as a setup gap (see Dockerfile yt-dlp install).
             "cannot import yt-dlp",
             "cannot import youtube-dl",
         ]
+        all_lines = combined.split('\n')
         error_lines = [
-                line for line in combined.split('\n')
+            line for line in all_lines
             if any(ind in line for ind in actual_error_indicators)
         ]
-            if error_lines and all(
+        per_item_error_lines = [
-                any(p in line for p in per_item_patterns)
+            line for line in error_lines
-                for line in error_lines
+            if any(p in line for p in per_item_patterns)
-            ):
+        ]
-                has_actual_error = False
+        source_level_error_lines = [
+            line for line in error_lines
+            if line not in set(per_item_error_lines)
+        ]
+
+        # Only source-level errors block skip/tier-limited classification.
+        # Per-item failures are tracked separately in run_stats.
+        has_actual_error = bool(source_level_error_lines)
+
+        # Strip per-item error lines from the text used for downstream pattern
+        # matching. Otherwise noise like "HTTP Error 403: Forbidden" from a
+        # yt-dlp HLS failure latches onto ACCESS_DENIED and masks the real
+        # signal (tier gating, no new content, etc.).
+        if per_item_error_lines:
+            per_item_set = set(per_item_error_lines)
+            combined = '\n'.join(line for line in all_lines if line not in per_item_set)
 
         # If we have skip evidence and no real errors, it's no_new_content
         if (skip_line_count > 0 or has_skip_text) and not has_actual_error:

backend/tests/services/test_gallery_dl_categorize.py

@@ -163,6 +163,36 @@ def test_skip_activity_takes_precedence_over_tier_limited():
     assert error_type == ErrorType.NO_NEW_CONTENT
 
 
+def test_tier_limited_with_ytdl_403s_not_access_denied():
+    """Real-world Patreon scenario: account lost tier access. Posts show up as
+    '[patreon][warning] Not allowed to view post N' (no downloads attempted)
+    AND yt-dlp tries to fetch video HLS manifests that are also tier-gated,
+    producing '[downloader.ytdl][error] HTTP Error 403: Forbidden' per item.
+
+    Before the fix, the 'Forbidden' text in per-item ytdl errors latched onto
+    ACCESS_DENIED_PATTERNS and misclassified the whole source as ACCESS_DENIED.
+    Correct behavior: TIER_LIMITED (the warnings are the dominant signal; the
+    per-item 403s are just tier gating manifesting at the HLS layer)."""
+    service = _make_service()
+    tier_warnings = "\n".join(
+        f"[patreon][warning] Not allowed to view post {100000 + i}"
+        for i in range(20)
+    )
+    ytdl_errors = "\n".join(
+        "[downloader.ytdl][error] ExtractorError: Failed to download m3u8 "
+        "information: HTTP Error 403: Forbidden "
+        f"(caused by <HTTPError 403: Forbidden>) ({i}/4)"
+        for i in range(1, 5)
+    ) + "\n[download][error] Failed to download 02_video.mp4\n"
+    stderr = tier_warnings + "\n" + ytdl_errors
+    stdout = ""
+
+    error_type, message = service._categorize_error(1, stdout, stderr)
+
+    assert error_type == ErrorType.TIER_LIMITED, f"got {error_type}: {message}"
+    assert "20" in message
+
+
 def test_ytdl_missing_with_skips_classifies_as_no_new_content():
     """Real-world scenario from a Patreon source: container is missing yt-dlp
     so every video logs '[downloader.ytdl][error] Cannot import yt-dlp' plus