session-memory: denoise error fingerprints (WP-0006 follow-up)

Tighten _is_failed: exclude successful hub JSON responses (top-level no-error
payloads) and file-read snapshots (numbered cat -n source lines) that were
polluting error_snippets. JSON verdict classifies error vs success payloads
directly. Cuts distinct fingerprints 444 -> 269 (~40%) over the real corpus with
the top errors unchanged. Assessment caveat updated. 5 new tests; suite 102/102.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

session_memory/core/digest.py

@@ -12,6 +12,7 @@ belongs to the Detect phase (PRD §6.2).
 from __future__ import annotations
 
 import collections
+import json
 import re
 from typing import Any
 
@@ -22,6 +23,12 @@ _FAIL_HINTS = ("error", "failed", "exception", "traceback", "fatal", "non-zero")
 # Substrings suggesting a clean test pass.
 _PASS_HINTS = ("passed", "0 failed", "ok", "success")
 
+# A line that is numbered source content from a Read result (`cat -n` style),
+# e.g. "229\t    raise InfospaceError(" — code text, never a runtime error.
+_NUMBERED_LINE_RE = re.compile(r"^\s*\d+\t")
+# Top-level keys that mark a JSON tool-result as an actual error (vs. success).
+_JSON_ERROR_KEYS = ("error", "errors", "detail")
+
 # Normalization patterns so the same error collapses to one fingerprint
 # regardless of paths / ids / counts (WP-0006 T01).
 _UUID_RE = re.compile(r"\b[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}\b", re.I)
@@ -195,12 +202,48 @@ def _error_body(event: SessionEvent, blobs: dict) -> str:
     return event.summary or ""
 
 
+def _looks_like_file_read(body: str) -> bool:
+    """True if the body is mostly numbered source lines (a Read result), not an error."""
+    lines = [ln for ln in body.splitlines() if ln.strip()]
+    if not lines:
+        return False
+    numbered = sum(1 for ln in lines if _NUMBERED_LINE_RE.match(ln))
+    return numbered >= max(3, len(lines) // 2)
+
+
+def _json_verdict(body: str):
+    """Classify a JSON tool-result body: 'error', 'success', or None (not JSON).
+
+    Hub MCP successes look like ``{"result": "..."}`` and mention 'error' deep
+    inside summaries but are not failures ('success'). A payload with a top-level
+    error key (``{"detail": ...}`` / ``{"error": ...}``) is 'error'. Non-JSON text
+    returns None so the plain fail-hint heuristic still applies.
+    """
+    s = body.strip()
+    if not s or s[0] not in "{[":
+        return None
+    try:
+        obj = json.loads(s)
+    except (ValueError, TypeError):
+        return None
+    if isinstance(obj, dict) and any(k in obj for k in _JSON_ERROR_KEYS):
+        return "error"
+    return "success"
+
+
 def _is_failed(event: SessionEvent, blobs: dict) -> bool:
     if event.kind == "error":
         return True
     if event.kind == "tool_result":
-        body = _error_body(event, blobs).lower()
-        return bool(body) and any(h in body for h in _FAIL_HINTS)
+        body = _error_body(event, blobs)
+        if not body.strip():
+            return False
+        if _looks_like_file_read(body):
+            return False
+        verdict = _json_verdict(body)
+        if verdict is not None:
+            return verdict == "error"
+        return any(h in body.lower() for h in _FAIL_HINTS)
     return False