diff --git a/regime/features.py b/regime/features.py
index 19b4f64..6c9df53 100644
--- a/regime/features.py
+++ b/regime/features.py
@@ -49,6 +49,20 @@
 DEFAULT_S3_BUCKET: str = "alpha-engine-research"
 DEFAULT_PRICE_CACHE_PREFIX: str = "predictor/price_cache/"
 
+# Wave-3 reader migration (ROADMAP L1401): producer write-both PR1
+# (alpha-engine-data#270, shipped 2026-05-19) seeded the new
+# ``reference/price_cache/`` prefix. During the ≥1-week soak window
+# (~2026-05-19 → 2026-05-26) every reader should consult the new
+# prefix first and fall back to legacy on miss. ``_read_parquet_close``
+# below iterates this list when the caller uses the legacy default;
+# explicit-prefix callers (tests, custom configs) opt out and get
+# single-prefix semantics. Wave-3 PR4 cutover drops the legacy entry
+# in a one-line edit here.
+_PRICE_CACHE_NEW_PREFIX: str = "reference/price_cache/"
+_PRICE_CACHE_FALLBACK_PREFIXES: tuple[str, ...] = (
+    _PRICE_CACHE_NEW_PREFIX, DEFAULT_PRICE_CACHE_PREFIX,
+)
+
 
 # Source tickers expected as ``{prefix}{ticker}.parquet`` keys. Index
 # symbols (VIX, VIX3M, TNX) require a leading caret in yfinance (^VIX,
@@ -69,14 +83,35 @@ def _read_parquet_close(
 ) -> pd.Series:
     """Read ``{prefix}{ticker}.parquet`` from S3 and return its Close
     column as a Series indexed by date. Returns an empty Series when
-    the key is missing — callers decide whether absence is fatal."""
-    key = f"{prefix}{ticker}.parquet"
-    try:
-        obj = s3_client.get_object(Bucket=bucket, Key=key)
-    except Exception as e:
+    the key is missing — callers decide whether absence is fatal.
+
+    When ``prefix == DEFAULT_PRICE_CACHE_PREFIX`` (the production
+    default), iterates the Wave-3 fallback list (new → legacy). When
+    ``prefix`` is anything else (test/override) we treat it as an
+    explicit single-prefix read.
+    """
+    # Pick the active read order. Default → fallback chain; custom →
+    # single-prefix.
+    if prefix == DEFAULT_PRICE_CACHE_PREFIX:
+        candidate_prefixes: tuple[str, ...] = _PRICE_CACHE_FALLBACK_PREFIXES
+    else:
+        candidate_prefixes = (prefix,)
+
+    obj = None
+    last_key: str | None = None
+    last_exc: Exception | None = None
+    for cand in candidate_prefixes:
+        last_key = f"{cand}{ticker}.parquet"
+        try:
+            obj = s3_client.get_object(Bucket=bucket, Key=last_key)
+            break
+        except Exception as e:
+            last_exc = e
+
+    if obj is None:
         logger.warning(
             "[regime_features] missing %s at s3://%s/%s (%s) — returning empty series",
-            ticker, bucket, key, type(e).__name__,
+            ticker, bucket, last_key, type(last_exc).__name__ if last_exc else "unknown",
         )
         return pd.Series(name=ticker, dtype="float64")
 
diff --git a/tests/test_regime_features.py b/tests/test_regime_features.py
index 623e977..5b87b6b 100644
--- a/tests/test_regime_features.py
+++ b/tests/test_regime_features.py
@@ -190,3 +190,77 @@ def test_source_tickers_match_data_side_convention() -> None:
     convention."""
     for ticker in SOURCE_TICKERS:
         assert not ticker.startswith("^"), f"{ticker} should not have ^ prefix"
+
+
+# ── Wave-3 reader migration (ROADMAP L1401) ─────────────────────────────────
+
+
+def test_wave3_reader_prefers_new_prefix_over_legacy() -> None:
+    """During the producer write-both soak (alpha-engine-data#270 shipped
+    2026-05-19; ≥1-week soak), both prefixes hold byte-equal copies. The
+    reader MUST consult ``reference/price_cache/`` first so the new home
+    is exercised end-to-end before PR4 deletes legacy."""
+    from regime.features import _PRICE_CACHE_NEW_PREFIX, _read_parquet_close
+
+    s3 = _FakeS3()
+    # Seed the NEW prefix with one Close value, LEGACY with a clearly-
+    # different one. If the reader picked legacy, we'd see 999.
+    new_df = pd.DataFrame(
+        {"Close": [100.0]},
+        index=pd.to_datetime(["2026-05-19"]),
+    )
+    legacy_df = pd.DataFrame(
+        {"Close": [999.0]},
+        index=pd.to_datetime(["2026-05-19"]),
+    )
+    s3.put_parquet("bkt", f"{_PRICE_CACHE_NEW_PREFIX}SPY.parquet", new_df)
+    s3.put_parquet("bkt", f"{DEFAULT_PRICE_CACHE_PREFIX}SPY.parquet", legacy_df)
+
+    s = _read_parquet_close(
+        "SPY", s3_client=s3, bucket="bkt",
+        prefix=DEFAULT_PRICE_CACHE_PREFIX,
+    )
+    assert s.iloc[0] == 100.0, "reader picked legacy when new was present"
+
+
+def test_wave3_reader_falls_back_to_legacy_when_new_missing() -> None:
+    """Early in the soak window the NEW prefix can lag a fresh ticker.
+    The reader must fall back to legacy gracefully."""
+    from regime.features import _read_parquet_close
+
+    s3 = _FakeS3()
+    legacy_df = pd.DataFrame(
+        {"Close": [123.0]},
+        index=pd.to_datetime(["2026-05-19"]),
+    )
+    # Only legacy seeded (no NEW key).
+    s3.put_parquet("bkt", f"{DEFAULT_PRICE_CACHE_PREFIX}SPY.parquet", legacy_df)
+
+    s = _read_parquet_close(
+        "SPY", s3_client=s3, bucket="bkt",
+        prefix=DEFAULT_PRICE_CACHE_PREFIX,
+    )
+    assert s.iloc[0] == 123.0
+
+
+def test_wave3_reader_explicit_custom_prefix_does_not_fall_back() -> None:
+    """Test/config-override callers that pass a non-default prefix opt
+    out of the fallback chain — single-prefix semantics. Mirrors the
+    write-side ``price_cache_write_prefixes`` opt-out."""
+    from regime.features import _read_parquet_close
+
+    s3 = _FakeS3()
+    # Seed only the legacy default — the custom prefix has nothing.
+    s3.put_parquet(
+        "bkt", f"{DEFAULT_PRICE_CACHE_PREFIX}SPY.parquet",
+        pd.DataFrame(
+            {"Close": [100.0]},
+            index=pd.to_datetime(["2026-05-19"]),
+        ),
+    )
+    s = _read_parquet_close(
+        "SPY", s3_client=s3, bucket="bkt",
+        prefix="custom/explicit/",  # NOT the default
+    )
+    # Custom prefix had no SPY.parquet → empty (no fallback to legacy).
+    assert s.empty