fix(hw-gate): keep Redline capture dumps out of the seat prompts

scripts/hw-gate/review.py

@@ -648,6 +648,27 @@ def build_prelim_prompt(select: dict, checkout: str, base: str, head: str, repo:
     return "\n".join(parts)
 
 
+# The diff is already capped at 400 KB below; the evidence had no cap, and the
+# first kernel-bucket run (#690, 33892920406) inlined ~2.5 MB of Redline
+# capture dumps: 1.36 M tokens against a 1 M window, both seats returned
+# nothing. run.py now elides the dumps at the source; this is the backstop
+# so no future field can push a seat prompt past the window again.
+EVIDENCE_PROMPT_CAP = 600 * 1024
+
+
+def evidence_for_prompt(evidence: dict | None) -> str:
+    if evidence is None:
+        return "(missing — hw-run did not produce evidence)"
+    text = json.dumps(evidence, indent=2, sort_keys=True)
+    raw = text.encode("utf-8")
+    if len(raw) <= EVIDENCE_PROMPT_CAP:
+        return text
+    return (
+        raw[:EVIDENCE_PROMPT_CAP].decode("utf-8", errors="ignore")
+        + f"\n\n[hw-gate.json truncated at {EVIDENCE_PROMPT_CAP} bytes of {len(raw)}; read the full file under the evidence directory]"
+    )
+
+
 def build_verdict_prompt(prelim_prompt: str, prelim: dict | None, evidence: dict | None, select: dict, hw_run_result: str) -> str:
     parts: list[str] = []
     parts.append(prelim_prompt)
@@ -657,11 +678,8 @@ def build_verdict_prompt(prelim_prompt: str, prelim: dict | None, evidence: dict
     parts.append("")
     parts.append(f"hw-run result: {hw_run_result}")
     parts.append("")
-    if evidence is not None:
-        parts.append("hw-gate.json:")
-        parts.append(json.dumps(evidence, indent=2, sort_keys=True))
-    else:
-        parts.append("hw-gate.json: (missing — hw-run did not produce evidence)")
+    parts.append("hw-gate.json:")
+    parts.append(evidence_for_prompt(evidence))
     return "\n".join(parts)
 
 
@@ -702,11 +720,8 @@ def build_decide_prompt(select: dict, prelim: dict | None, evidence: dict | None
     parts.append("Sol prelim (prelim.json prelim field):")
     parts.append(json.dumps(prelim, indent=2, sort_keys=True) if prelim is not None else "null")
     parts.append("")
-    if evidence is not None:
-        parts.append("hw-gate.json evidence:")
-        parts.append(json.dumps(evidence, indent=2, sort_keys=True))
-    else:
-        parts.append("hw-gate.json: missing")
+    parts.append("hw-gate.json evidence:")
+    parts.append(evidence_for_prompt(evidence))
     parts.append("")
     parts.append("Sol verdict (verdict.json verdict+floor):")
     parts.append(json.dumps(sol_verdict_obj, indent=2, sort_keys=True) if sol_verdict_obj is not None else "null")

scripts/hw-gate/run.py

@@ -773,7 +773,36 @@ def run_kernel(repo, models_dir, kernel_cfg, fixtures_manifest, env, logs_dir) -
         else:
             reason = "no boolean parity summary in report (fail closed)"
             status = "fail"
-    return {"report": report, "exit": exit_code, "stderr_tail": stderr[-2000:] if stderr else "", "status": status, "reason": reason}
+    return {"report": elide_captures(report), "exit": exit_code, "stderr_tail": stderr[-2000:] if stderr else "", "status": status, "reason": reason}
+
+
+# Keys under which redline_daemon_harness.py stores raw per-dispatch capture
+# dumps. They are what makes the report large (516 KB per lane on #690's
+# first kernel-bucket run) and carry nothing a reviewer reads: the verdict
+# fields (`pass`, `sequence_stable`, `measurement`, `aql_shadow`, failures)
+# stay. The full report is still on disk as hw-gate-logs/redline.json.
+_CAPTURE_KEYS = {"captures", "dispatches", "packets", "raw"}
+
+
+def elide_captures(report):
+    """Return `report` with capture dumps replaced by their size, recursively.
+
+    Both seat prompts embed the merged evidence JSON verbatim; with the dumps
+    in, #690's decide prompt was 1.36 M tokens against a 1 M window and both
+    seats returned nothing (run 33892920406).
+    """
+    if isinstance(report, dict):
+        out = {}
+        for key, value in report.items():
+            if key in _CAPTURE_KEYS and isinstance(value, (list, dict)) and len(json.dumps(value)) > 4096:
+                n = len(value)
+                out[key] = f"<{n} entries elided; full report in hw-gate-logs/redline.json>"
+            else:
+                out[key] = elide_captures(value)
+        return out
+    if isinstance(report, list):
+        return [elide_captures(v) for v in report]
+    return report
 
 
 def render_md(evidence: dict) -> str: