diff --git a/tests/test_benchmark_matrix.py b/tests/test_benchmark_matrix.py
index c6844621..8915cba6 100644
--- a/tests/test_benchmark_matrix.py
+++ b/tests/test_benchmark_matrix.py
@@ -1,9 +1,11 @@
 """Unit tests for the reuse x context benchmark-matrix harness (pure-logic core).
 
-Covers the platform-independent modules: corpus construction, paired statistics,
-cache-hit measurement/classification, memory/OOM classification, thermal math,
-matrix spec, reporting, host parsing, and cell-level pure helpers. The bench-host
-modules (systems/cell orchestration) are exercised on Apple Silicon, not here.
+Covers prompt construction and paired statistics: corpus construction, paired
+statistics, and cache-hit measurement/classification. Memory/OOM classification,
+thermal math, matrix spec, reporting, host parsing, and cell-level pure helpers
+live in test_benchmark_matrix_reporting.py — split for the 500-line file gate.
+The bench-host modules (systems/cell orchestration) are exercised on Apple
+Silicon, not here.
 """
 
 from __future__ import annotations
@@ -13,18 +15,7 @@
 
 import pytest
 
-from benchmarks.ollama_vs_squish.matrix import (
-    cache_probe,
-    corpus,
-    host,
-    matrix_spec,
-    memory,
-    report,
-    stats_ext,
-    systems,
-    thermal,
-)
-from benchmarks.ollama_vs_squish.matrix.cell import _paired, inter_token_stats
+from benchmarks.ollama_vs_squish.matrix import cache_probe, corpus, stats_ext, systems
 
 
 # ── mock tokenizer ────────────────────────────────────────────────────────────
@@ -295,249 +286,3 @@ def test_classify_exact_and_unknown_nan():
     assert cache_probe.classify("squish", 1.0, 0.99, "m").status == "ok"
     v = cache_probe.classify("squish", 0.5, float("nan"), "radix_only")
     assert v.status == "unknown" and v.measured is None
-
-
-# ── memory ────────────────────────────────────────────────────────────────────
-
-GB = 1024**3
-
-
-def test_memory_fit():
-    ms = memory.classify_memory_status(
-        peak_rss_bytes=8 * GB, ram_bytes=16 * GB, request_failed=False
-    )
-    assert ms.status == "fit" and ms.fit
-
-
-def test_memory_oom_on_failure():
-    ms = memory.classify_memory_status(
-        peak_rss_bytes=16 * GB, ram_bytes=16 * GB, request_failed=True, oom_signal=True
-    )
-    assert ms.status == "oom"
-
-
-def test_memory_degraded_near_cap():
-    ms = memory.classify_memory_status(
-        peak_rss_bytes=int(15.8 * GB), ram_bytes=16 * GB, request_failed=False
-    )
-    assert ms.status == "degraded_via_governor"
-
-
-def test_memory_degraded_throughput_collapse():
-    ms = memory.classify_memory_status(
-        peak_rss_bytes=8 * GB,
-        ram_bytes=16 * GB,
-        request_failed=False,
-        decode_tps=5.0,
-        baseline_tps=40.0,
-    )
-    assert ms.status == "degraded_via_governor"
-
-
-def test_scan_log_for_signals():
-    oom, gov = memory.scan_log_for_signals("... Out of memory: failed to allocate ...")
-    assert oom is True
-    _, gov2 = memory.scan_log_for_signals("memory pressure detected, compressed pages")
-    assert gov2 is True
-
-
-def test_kv_cache_mb_from_metrics():
-    assert memory.kv_cache_mb_from_metrics({"squish_kv_cache_memory_mb": 42.0}) == 42.0
-    assert memory.kv_cache_mb_from_metrics({}) is None
-
-
-# ── thermal ───────────────────────────────────────────────────────────────────
-
-
-def test_parse_powermetrics_temp():
-    text = "CPU die temperature: 47.8 C\nother: 1"
-    assert thermal.parse_powermetrics_temp(text) == 47.8
-
-
-def test_parse_macmon_temp_takes_max_of_cpu_gpu():
-    line = '{"temp": {"cpu_temp_avg": 61.5, "gpu_temp_avg": 68.2}}'
-    assert thermal.parse_macmon_temp(line) == 68.2
-
-
-def test_parse_macmon_temp_uses_last_line_of_stream():
-    text = '{"temp": {"cpu_temp_avg": 40.0}}\n{"temp": {"cpu_temp_avg": 50.0}}'
-    assert thermal.parse_macmon_temp(text) == 50.0
-
-
-def test_parse_macmon_temp_malformed_or_empty_returns_none():
-    assert thermal.parse_macmon_temp("not json") is None
-    assert thermal.parse_macmon_temp("") is None
-    assert thermal.parse_macmon_temp('{"temp": {}}') is None
-
-
-def test_plausible_die_temp_rejects_bogus_zero():
-    # osx-cpu-temp on Apple Silicon reads Intel-only SMC keys and returns a
-    # fixed 0.0 rather than failing — must be rejected, not trusted.
-    assert thermal._plausible_die_temp(0.0) is False
-    assert thermal._plausible_die_temp(47.8) is True
-    assert thermal._plausible_die_temp(200.0) is False
-
-
-def test_drift_check_pass_and_fail():
-    assert thermal.drift_check(20.0, 20.2).passed  # +1.0% <= 1.7%
-    assert not thermal.drift_check(20.0, 21.0).passed  # +5.0%
-    assert thermal.drift_check(0.0, 0.0).passed
-
-
-def test_wait_for_baseline_no_sensor_skips():
-    assert thermal.wait_for_baseline(reader=lambda: None, log=lambda *_: None) is True
-
-
-def test_wait_for_baseline_reaches_target():
-    seq = iter([70.0, 60.0, 49.0])
-    ok = thermal.wait_for_baseline(
-        target_c=50.0, poll_s=0.0, reader=lambda: next(seq), log=lambda *_: None
-    )
-    assert ok is True
-
-
-def test_thermal_log_records_and_max():
-    tl = thermal.ThermalLog()
-    tl.record(40.0, "a")
-    tl.record(None, "b")  # dropped
-    tl.record(55.0, "c")
-    assert tl.max_temp() == 55.0 and len(tl.as_rows()) == 2
-
-
-# ── matrix_spec ───────────────────────────────────────────────────────────────
-
-
-def test_all_cells_count():
-    cells = matrix_spec.all_cells()
-    assert len(cells) == len(matrix_spec.REUSE_LEVELS) * len(matrix_spec.CONTEXT_LENGTHS)
-    assert cells[0].cell_id == "r000_c4000"
-
-
-def test_kill_test_cell():
-    c = matrix_spec.kill_test_cell()
-    assert c.reuse == 0.5 and c.ctx_tokens == 8000
-
-
-def test_counterbalanced_order_rotates():
-    names = ["a", "b", "c"]
-    assert matrix_spec.counterbalanced_order(names, 0) == ["a", "b", "c"]
-    assert matrix_spec.counterbalanced_order(names, 1) == ["b", "c", "a"]
-    assert matrix_spec.counterbalanced_order([], 3) == []
-
-
-def test_second_model_rows():
-    rows = matrix_spec.second_model_rows()
-    assert all(r.reuse in (0.0, 0.5) for r in rows)
-
-
-# ── host ──────────────────────────────────────────────────────────────────────
-
-
-def test_parse_sysctl_memsize():
-    assert host.parse_sysctl_memsize("hw.memsize: 17179869184") == 17179869184
-    assert host.parse_sysctl_memsize("hw.memsize: ") is None
-
-
-# ── cell pure helpers ─────────────────────────────────────────────────────────
-
-
-def test_inter_token_stats():
-    stamps = [0.0, 1.0, 1.1, 1.2, 1.3]  # first gap (TTFT) excluded
-    s = inter_token_stats(stamps)
-    assert s["itl_count"] == 3 and s["itl_p50_ms"] is not None
-
-
-def test_inter_token_stats_too_few():
-    assert inter_token_stats([0.0, 1.0])["itl_p50_ms"] is None
-
-
-class _Run:
-    def __init__(self, i, val, failed=False):
-        self.run_index = i
-        self.e2e_s = val
-        self.failed = failed
-
-
-class _Sys:
-    def __init__(self, runs):
-        self.runs = runs
-
-
-def test_paired_aligns_common_indices():
-    sa = _Sys([_Run(0, 1.0), _Run(1, 2.0), _Run(2, None)])
-    sb = _Sys([_Run(0, 3.0), _Run(1, 4.0, failed=True), _Run(2, 5.0)])
-    va, vb = _paired(sa, sb, "e2e_s")
-    assert va == [1.0] and vb == [3.0]  # only index 0 valid in both
-
-
-# ── report ────────────────────────────────────────────────────────────────────
-
-
-def _fake_cell(reuse, ctx, squish_med, ollama_med, p, delta):
-    return {
-        "cell_id": f"r{int(reuse * 100):03d}_c{ctx}",
-        "reuse": reuse,
-        "ctx_tokens": ctx,
-        "status": "ok",
-        "systems": {
-            "squish_int4": {
-                "role": "head_to_head",
-                "runs": [{"failed": False}] * 30,
-                "cache_total_runs": 30,
-                "cache_ok_runs": 30,
-            },
-            "ollama_q4km": {
-                "role": "head_to_head",
-                "runs": [{"failed": False}] * 30,
-                "cache_total_runs": 30,
-                "cache_ok_runs": 30,
-            },
-        },
-        "comparisons": {
-            "e2e_s": {
-                "a_system": "squish_int4",
-                "b_system": "ollama_q4km",
-                "a": {"median": squish_med, "iqr": 0.1},
-                "b": {"median": ollama_med, "iqr": 0.2},
-                "wilcoxon": {"p_value": p},
-                "cliffs_delta": delta,
-                "cliffs_magnitude": "large",
-            }
-        },
-    }
-
-
-def test_speed_ratio_lower_is_better():
-    cell = _fake_cell(0.0, 8000, squish_med=1.0, ollama_med=2.0, p=0.001, delta=-1.0)
-    # e2e lower is better: ratio = ollama/squish = 2.0
-    assert report.speed_ratio(cell, "e2e_s") == 2.0
-
-
-def test_metric_table_contains_cells():
-    cells = [_fake_cell(0.0, 8000, 1.0, 2.0, 0.001, -1.0)]
-    table = report.metric_table(cells, "e2e_s", [0.0], [8000])
-    assert "End-to-end" in table and "p=0.0010" in table
-
-
-def test_one_screen_summary_headlines():
-    cells = [
-        _fake_cell(0.0, 8000, 1.0, 2.0, 0.001, -1.0),
-        _fake_cell(1.0, 8000, 0.5, 2.0, 0.0001, -1.0),
-    ]
-    out = report.one_screen_summary(cells, [8000])
-    assert "ONE-SCREEN SUMMARY" in out
-    assert "Cold / unique" in out and "Exact-repeat" in out
-
-
-def test_postflight_pass():
-    cells = [_fake_cell(0.0, 8000, 1.0, 2.0, 0.001, -1.0)]
-    out = report.postflight(cells, min_runs=30)
-    assert "POST-FLIGHT VERIFICATION" in out
-    assert "[FAIL]" not in out
-
-
-def test_postflight_flags_cache_mismatch():
-    cell = _fake_cell(0.0, 8000, 1.0, 2.0, 0.001, -1.0)
-    cell["systems"]["squish_int4"]["cache_ok_runs"] = 10  # mismatch
-    out = report.postflight([cell], min_runs=30)
-    assert "[FAIL]" in out
diff --git a/tests/test_benchmark_matrix_reporting.py b/tests/test_benchmark_matrix_reporting.py
new file mode 100644
index 00000000..120923b1
--- /dev/null
+++ b/tests/test_benchmark_matrix_reporting.py
@@ -0,0 +1,258 @@
+"""Unit tests for the benchmark-matrix harness — memory/thermal/reporting half.
+
+Split out of test_benchmark_matrix.py (which covers corpus construction, paired
+statistics, and cache-hit measurement/classification) to keep both files under
+the 500-line gate. Covers: memory/OOM classification, thermal math, matrix
+spec, host parsing, cell-level pure helpers, and reporting. The bench-host
+modules (systems/cell orchestration) are exercised on Apple Silicon, not here.
+"""
+
+from __future__ import annotations
+
+from benchmarks.ollama_vs_squish.matrix import host, matrix_spec, memory, report, thermal
+from benchmarks.ollama_vs_squish.matrix.cell import _paired, inter_token_stats
+
+# ── memory ────────────────────────────────────────────────────────────────────
+
+GB = 1024**3
+
+
+def test_memory_fit():
+    ms = memory.classify_memory_status(
+        peak_rss_bytes=8 * GB, ram_bytes=16 * GB, request_failed=False
+    )
+    assert ms.status == "fit" and ms.fit
+
+
+def test_memory_oom_on_failure():
+    ms = memory.classify_memory_status(
+        peak_rss_bytes=16 * GB, ram_bytes=16 * GB, request_failed=True, oom_signal=True
+    )
+    assert ms.status == "oom"
+
+
+def test_memory_degraded_near_cap():
+    ms = memory.classify_memory_status(
+        peak_rss_bytes=int(15.8 * GB), ram_bytes=16 * GB, request_failed=False
+    )
+    assert ms.status == "degraded_via_governor"
+
+
+def test_memory_degraded_throughput_collapse():
+    ms = memory.classify_memory_status(
+        peak_rss_bytes=8 * GB,
+        ram_bytes=16 * GB,
+        request_failed=False,
+        decode_tps=5.0,
+        baseline_tps=40.0,
+    )
+    assert ms.status == "degraded_via_governor"
+
+
+def test_scan_log_for_signals():
+    oom, gov = memory.scan_log_for_signals("... Out of memory: failed to allocate ...")
+    assert oom is True
+    _, gov2 = memory.scan_log_for_signals("memory pressure detected, compressed pages")
+    assert gov2 is True
+
+
+def test_kv_cache_mb_from_metrics():
+    assert memory.kv_cache_mb_from_metrics({"squish_kv_cache_memory_mb": 42.0}) == 42.0
+    assert memory.kv_cache_mb_from_metrics({}) is None
+
+
+# ── thermal ───────────────────────────────────────────────────────────────────
+
+
+def test_parse_powermetrics_temp():
+    text = "CPU die temperature: 47.8 C\nother: 1"
+    assert thermal.parse_powermetrics_temp(text) == 47.8
+
+
+def test_parse_macmon_temp_takes_max_of_cpu_gpu():
+    line = '{"temp": {"cpu_temp_avg": 61.5, "gpu_temp_avg": 68.2}}'
+    assert thermal.parse_macmon_temp(line) == 68.2
+
+
+def test_parse_macmon_temp_uses_last_line_of_stream():
+    text = '{"temp": {"cpu_temp_avg": 40.0}}\n{"temp": {"cpu_temp_avg": 50.0}}'
+    assert thermal.parse_macmon_temp(text) == 50.0
+
+
+def test_parse_macmon_temp_malformed_or_empty_returns_none():
+    assert thermal.parse_macmon_temp("not json") is None
+    assert thermal.parse_macmon_temp("") is None
+    assert thermal.parse_macmon_temp('{"temp": {}}') is None
+
+
+def test_plausible_die_temp_rejects_bogus_zero():
+    # osx-cpu-temp on Apple Silicon reads Intel-only SMC keys and returns a
+    # fixed 0.0 rather than failing — must be rejected, not trusted.
+    assert thermal._plausible_die_temp(0.0) is False
+    assert thermal._plausible_die_temp(47.8) is True
+    assert thermal._plausible_die_temp(200.0) is False
+
+
+def test_drift_check_pass_and_fail():
+    assert thermal.drift_check(20.0, 20.2).passed  # +1.0% <= 1.7%
+    assert not thermal.drift_check(20.0, 21.0).passed  # +5.0%
+    assert thermal.drift_check(0.0, 0.0).passed
+
+
+def test_wait_for_baseline_no_sensor_skips():
+    assert thermal.wait_for_baseline(reader=lambda: None, log=lambda *_: None) is True
+
+
+def test_wait_for_baseline_reaches_target():
+    seq = iter([70.0, 60.0, 49.0])
+    ok = thermal.wait_for_baseline(
+        target_c=50.0, poll_s=0.0, reader=lambda: next(seq), log=lambda *_: None
+    )
+    assert ok is True
+
+
+def test_thermal_log_records_and_max():
+    tl = thermal.ThermalLog()
+    tl.record(40.0, "a")
+    tl.record(None, "b")  # dropped
+    tl.record(55.0, "c")
+    assert tl.max_temp() == 55.0 and len(tl.as_rows()) == 2
+
+
+# ── matrix_spec ───────────────────────────────────────────────────────────────
+
+
+def test_all_cells_count():
+    cells = matrix_spec.all_cells()
+    assert len(cells) == len(matrix_spec.REUSE_LEVELS) * len(matrix_spec.CONTEXT_LENGTHS)
+    assert cells[0].cell_id == "r000_c4000"
+
+
+def test_kill_test_cell():
+    c = matrix_spec.kill_test_cell()
+    assert c.reuse == 0.5 and c.ctx_tokens == 8000
+
+
+def test_counterbalanced_order_rotates():
+    names = ["a", "b", "c"]
+    assert matrix_spec.counterbalanced_order(names, 0) == ["a", "b", "c"]
+    assert matrix_spec.counterbalanced_order(names, 1) == ["b", "c", "a"]
+    assert matrix_spec.counterbalanced_order([], 3) == []
+
+
+def test_second_model_rows():
+    rows = matrix_spec.second_model_rows()
+    assert all(r.reuse in (0.0, 0.5) for r in rows)
+
+
+# ── host ──────────────────────────────────────────────────────────────────────
+
+
+def test_parse_sysctl_memsize():
+    assert host.parse_sysctl_memsize("hw.memsize: 17179869184") == 17179869184
+    assert host.parse_sysctl_memsize("hw.memsize: ") is None
+
+
+# ── cell pure helpers ─────────────────────────────────────────────────────────
+
+
+def test_inter_token_stats():
+    stamps = [0.0, 1.0, 1.1, 1.2, 1.3]  # first gap (TTFT) excluded
+    s = inter_token_stats(stamps)
+    assert s["itl_count"] == 3 and s["itl_p50_ms"] is not None
+
+
+def test_inter_token_stats_too_few():
+    assert inter_token_stats([0.0, 1.0])["itl_p50_ms"] is None
+
+
+class _Run:
+    def __init__(self, i, val, failed=False):
+        self.run_index = i
+        self.e2e_s = val
+        self.failed = failed
+
+
+class _Sys:
+    def __init__(self, runs):
+        self.runs = runs
+
+
+def test_paired_aligns_common_indices():
+    sa = _Sys([_Run(0, 1.0), _Run(1, 2.0), _Run(2, None)])
+    sb = _Sys([_Run(0, 3.0), _Run(1, 4.0, failed=True), _Run(2, 5.0)])
+    va, vb = _paired(sa, sb, "e2e_s")
+    assert va == [1.0] and vb == [3.0]  # only index 0 valid in both
+
+
+# ── report ────────────────────────────────────────────────────────────────────
+
+
+def _fake_cell(reuse, ctx, squish_med, ollama_med, p, delta):
+    return {
+        "cell_id": f"r{int(reuse * 100):03d}_c{ctx}",
+        "reuse": reuse,
+        "ctx_tokens": ctx,
+        "status": "ok",
+        "systems": {
+            "squish_int4": {
+                "role": "head_to_head",
+                "runs": [{"failed": False}] * 30,
+                "cache_total_runs": 30,
+                "cache_ok_runs": 30,
+            },
+            "ollama_q4km": {
+                "role": "head_to_head",
+                "runs": [{"failed": False}] * 30,
+                "cache_total_runs": 30,
+                "cache_ok_runs": 30,
+            },
+        },
+        "comparisons": {
+            "e2e_s": {
+                "a_system": "squish_int4",
+                "b_system": "ollama_q4km",
+                "a": {"median": squish_med, "iqr": 0.1},
+                "b": {"median": ollama_med, "iqr": 0.2},
+                "wilcoxon": {"p_value": p},
+                "cliffs_delta": delta,
+                "cliffs_magnitude": "large",
+            }
+        },
+    }
+
+
+def test_speed_ratio_lower_is_better():
+    cell = _fake_cell(0.0, 8000, squish_med=1.0, ollama_med=2.0, p=0.001, delta=-1.0)
+    # e2e lower is better: ratio = ollama/squish = 2.0
+    assert report.speed_ratio(cell, "e2e_s") == 2.0
+
+
+def test_metric_table_contains_cells():
+    cells = [_fake_cell(0.0, 8000, 1.0, 2.0, 0.001, -1.0)]
+    table = report.metric_table(cells, "e2e_s", [0.0], [8000])
+    assert "End-to-end" in table and "p=0.0010" in table
+
+
+def test_one_screen_summary_headlines():
+    cells = [
+        _fake_cell(0.0, 8000, 1.0, 2.0, 0.001, -1.0),
+        _fake_cell(1.0, 8000, 0.5, 2.0, 0.0001, -1.0),
+    ]
+    out = report.one_screen_summary(cells, [8000])
+    assert "ONE-SCREEN SUMMARY" in out
+    assert "Cold / unique" in out and "Exact-repeat" in out
+
+
+def test_postflight_pass():
+    cells = [_fake_cell(0.0, 8000, 1.0, 2.0, 0.001, -1.0)]
+    out = report.postflight(cells, min_runs=30)
+    assert "POST-FLIGHT VERIFICATION" in out
+    assert "[FAIL]" not in out
+
+
+def test_postflight_flags_cache_mismatch():
+    cell = _fake_cell(0.0, 8000, 1.0, 2.0, 0.001, -1.0)
+    cell["systems"]["squish_int4"]["cache_ok_runs"] = 10  # mismatch
+    out = report.postflight([cell], min_runs=30)
+    assert "[FAIL]" in out