test(v7-h41): sustained gen.run does not leak Metal allocator peak

Closes plan item #41. 5 × max_new_tokens=1024 + kv_cache_layers=2
through gen.run, assert peak after run 5 < 1.2× peak after run 1.
Warm-up + reset_peak before measurement so import-time allocator
init doesn't bias the first reading. Skips cleanly on non-Metal
backends or when peak reports 0.

1/1 pytest.

Author: apstenku123

tests/v4/test_long_gen_memory_bounded.py

@@ -0,0 +1,73 @@
+"""V7-H41: long-form sustained generation does not leak memory.
+
+Pins that running gen.run repeatedly with max_new_tokens=1024 doesn't
+balloon Metal allocator peak across invocations. Each gen.run builds
++ tears down its own KVCache, samplers, and event-bus subscribers;
+if any of those leak references the peak memory after 5 runs would
+be N× larger than after the first.
+
+Acceptance: peak_memory after run 5 < 1.2× peak after run 1.
+"""
+
+from __future__ import annotations
+
+import pytest
+
+import mlx.core as mx
+
+from cppmega_v4.jsonrpc.gen_run_method import GenRunParams, gen_run
+
+
+def _has_metal_peak() -> bool:
+    return hasattr(mx, "get_peak_memory") or hasattr(mx, "metal")
+
+
+def _peak_bytes() -> int:
+    """Read Metal allocator peak in a backward-compatible way."""
+    if hasattr(mx, "get_peak_memory"):
+        return int(mx.get_peak_memory())
+    if hasattr(mx, "metal") and hasattr(mx.metal, "get_peak_memory"):
+        return int(mx.metal.get_peak_memory())
+    return 0
+
+
+def _reset_peak() -> None:
+    if hasattr(mx, "reset_peak_memory"):
+        mx.reset_peak_memory()
+    elif hasattr(mx, "metal") and hasattr(mx.metal, "reset_peak_memory"):
+        mx.metal.reset_peak_memory()
+
+
+@pytest.fixture(autouse=True)
+def _skip_without_metal():
+    if not _has_metal_peak():
+        pytest.skip("no Metal allocator peak — cannot bound memory")
+
+
+def _one_long_gen() -> None:
+    gen_run(GenRunParams(
+        prompt_tokens=[1], eos_token_id=99_999,
+        max_new_tokens=1024, strategy="greedy", seed=11,
+        kv_cache_layers=2, kv_cache_head_dim=16,
+    ))
+
+
+def test_v7_h41_repeated_long_gen_memory_bounded():
+    """5 × max_new_tokens=1024: peak after 5 < 1.2× peak after 1."""
+    # Warm up so the first measurement isn't dominated by import-time
+    # allocator initialisation.
+    _one_long_gen()
+    _reset_peak()
+    _one_long_gen()
+    peak_after_1 = _peak_bytes()
+    for _ in range(4):
+        _one_long_gen()
+    peak_after_5 = _peak_bytes()
+    if peak_after_1 == 0:
+        # Some backends report 0 even with metal present — skip cleanly.
+        pytest.skip("Metal peak reported 0 — cannot bound delta")
+    ratio = peak_after_5 / peak_after_1
+    assert ratio < 1.2, (
+        f"sustained gen leak: peak_after_1={peak_after_1} bytes, "
+        f"peak_after_5={peak_after_5} bytes, ratio={ratio:.3f}× "
+        f"(limit 1.2×)")