grinder: auto-dispatch @reviewer on escalation/regression/large diff

W31 self-improvement gap #2 (reviewer-confirmed): 125 GREENs shipped, only
41 ever reviewed. Most grinds are low-risk and fine to skip, but a GREEN
that escalated past the floor tier, had a regression along the way, or grew
past a minimal-fix size is exactly the kind of result likely to hide a real
problem in code nobody looks at again. grind()'s report() now emits a
machine-parseable "AUTO-REVIEW: yes (reasons)" marker on those GREENs;
serve() parses it and fires the same @reviewer ask a human would send,
reusing the existing wake-on-address mechanism (reviewer_lurker/CLAUDE.md)
rather than adding a new dispatch path.

bin/bullpen-grinder

@@ -31,6 +31,7 @@ TIERS  = os.environ.get("GRIND_TIERS", "deepseek-v4-flash-dspark").split(",")  #
 PY_TIERS = os.environ.get("PY_GRIND_TIERS", "deepseek-v4-flash-dspark,[free] qwen/qwen3-coder")
 MAX_ITERS   = 10          # total model attempts across all tiers
 ESCALATE_AT = 4           # attempts on a tier without progress before escalating
+LARGE_DIFF_LINES = 40     # insertions+deletions above which a GREEN still gets auto-flagged for review
 COLLECTION_PENALTY = 500  # a module that won't import scores worse than any assertion-fail count
 PYTEST_TIMEOUT = 120
 GO_TEST_TIMEOUT = 600     # compile + test is slower than pytest, especially cold
@@ -286,6 +287,13 @@ def _eval_escalation_discipline(rid, escalations, passed, final_model):
              f"final_model={final_model} passed={passed} note=\"{note}\" ts={int(time.time())}",
              ns="/trajectory/grinder")
 
+def _diffstat_lines(stat):
+    """Sum insertions+deletions out of a `git diff --stat` summary line
+    ('N files changed, A insertions(+), D deletions(-)'). 0 if unparsable."""
+    ins = re.search(r"(\d+) insertion", stat)
+    dels = re.search(r"(\d+) deletion", stat)
+    return (int(ins.group(1)) if ins else 0) + (int(dels.group(1)) if dels else 0)
+
 def grind(task, test_rel, targets):
     rid = str(int(time.time()))
     work = f"/tmp/grind-{rid}"
@@ -381,8 +389,22 @@ def grind(task, test_rel, targets):
             diff = sh(["git", "-C", work, "diff", baseline, b["sha"], "--", ".", *NOISE],
                       timeout=60).stdout
             if b["passed"]:
+                # #2 of the W31 self-improvement findings: 125 GREENs shipped, only 41 ever
+                # reviewed. Flag the risky subset for an automatic @reviewer dispatch (serve()
+                # below parses this line) instead of leaving every GREEN's review to chance —
+                # escalated past the floor tier, a regression happened along the way, or the
+                # diff grew past a "minimal fix" size are exactly the cases most likely to hide
+                # a real problem in code nobody looks at again.
+                escalated = b["model"] != TIERS[0]
+                n_regressions = sum(1 for e in escalations if e.get("reason") == "regression")
+                difflines = _diffstat_lines(stat)
+                reasons = []
+                if escalated: reasons.append(f"escalated to {b['model']}")
+                if n_regressions: reasons.append(f"{n_regressions} regression(s)")
+                if difflines > LARGE_DIFF_LINES: reasons.append(f"large diff ({difflines} lines)")
+                auto_review = ("AUTO-REVIEW: yes (" + ", ".join(reasons) + ")") if reasons else "AUTO-REVIEW: no"
                 return (f"GREEN ✅ — {test_rel} passes ({_model_tag(b)}).\n{stat}\n"
-                        f"worktree: {work} (diff below, apply after review)\n\n{diff[:3500]}")
+                        f"worktree: {work} (diff below, apply after review)\n{auto_review}\n\n{diff[:3500]}")
             return (f"RED ❌ — best-so-far {b['failed']} failing ({_model_tag(b)}); could not reach "
                     f"green in {MAX_ITERS} attempts. worktree left at {work} for a human.\n{stat}\n\n"
                     f"{diff[:3000]}\nlast pytest:\n{b['tail']}")
@@ -678,6 +700,11 @@ def _room_read(since):
             except Exception: pass
     return out
 
+def _auto_review_reason(reply):
+    """Pulls the reason string out of grind()'s 'AUTO-REVIEW: yes (...)' marker line, or None."""
+    m = re.search(r"^AUTO-REVIEW: yes \((.*)\)$", reply, re.M)
+    return m.group(1) if m else None
+
 def serve(nick):
     state = os.path.expanduser(f"~/.grinder-{nick}.since")
     def persist(v):
@@ -732,12 +759,16 @@ def serve(nick):
                     targets = [t for t in re.findall(r"\b[\w-]+(?:/[\w.-]+)*/[\w.-]+\.\w+\b", body)
                                if not t.startswith("tests/") and t != test_rel]
                     reply = _remote_grind(body, test_rel, targets)
+                review_reason = _auto_review_reason(reply)
                 posted = False
                 for _ in range(3):
                     if say(reply, to=f"@{asker}", typ="reply", rid=mid): posted = True; break
                     time.sleep(2)
                 if posted: since = mid; persist(since)
                 else: sys.stderr.write(f"@{nick}: reply failed id {mid}; retry\n"); break
+                if review_reason:
+                    say(f"🔍 auto-review — @{nick}'s grind for @{asker} flagged itself ({review_reason}); "
+                        f"please review before it ships.\n\n{reply}", to="@reviewer", typ="ask")
         except Exception as e:
             sys.stderr.write(f"serve loop error: {e}\n")
         time.sleep(3)