Add minute bar feature pipeline

docs/minute_bar_data.md

@@ -0,0 +1,67 @@
+# Minute Bar Data Notes
+
+The minute-bar path downloads raw Baostock intraday bars and stores them as a
+Hive-partitioned dataset:
+
+```bash
+uv run python cli.py data download-minute \
+    --universe sh600000 \
+    --start-date 2024-01-02 --end-date 2024-01-05 \
+    --frequency 5
+```
+
+The default layout is:
+
+```text
+data/minute_bars/{universe}/frequency=5m/month=YYYY-MM/*.pq
+```
+
+Feature plugins can aggregate those bars to daily `symbol_id,date` feature
+files, for example:
+
+```bash
+uv run python cli.py feature compute \
+    --minute-path data/minute_bars/sh600000 \
+    --daily-path data/daily_bars/sh600000 \
+    --feature-type minute_daily_summary \
+    --feature-name minute_summary
+```
+
+## Daily vs Minute Reconciliation
+
+Baostock's daily raw bars and 5-minute raw bars are close, but they should not
+be treated as perfectly reconstructible from each other.
+
+When checking consistency, compare daily raw bars (`data download --adjust none`)
+against minute bars on the same raw price scale. The minute aggregation should
+use:
+
+- `open`: first minute open
+- `high`: max minute high
+- `low`: min minute low
+- `close`: last minute close
+- `volume`: sum minute volume
+- `amount`: sum minute amount
+- `vwap`: `sum(amount) / sum(volume)`
+
+In a sanity check for `sh600000` from `2024-01-02` through `2024-01-05`, Baostock
+returned 4 daily rows and 192 5-minute bars, exactly 48 bars per day. Open, low,
+and close matched daily exactly on all 4 days. High matched on 3 of 4 days; on
+`2024-01-04`, the daily high was `6.67` while the max 5-minute high was `6.66`.
+Minute-summed volume and amount were higher than daily by roughly `0.16%` to
+`1.23%`. VWAP remained very close, with max relative difference around
+`0.0043%`.
+
+This appears to be a source-level Baostock reconciliation caveat, not a parser
+or ordering issue: the minute bars covered the regular `09:35:00` through
+`15:00:00` range and sorted correctly by timestamp.
+
+Practical guidance:
+
+- Use tolerance-based daily-vs-minute checks; do not require exact equality for
+  high, volume, or amount.
+- Expect open/close alignment to be a stronger sanity check than exact volume
+  reconstruction.
+- Use minute-derived values as separate daily features, not as replacements for
+  the canonical daily bar dataset unless a strategy explicitly wants that
+  source convention.