The 20 EMA: Only One Day in 16 Years Never Touched It

I have a CSV of cleaned 1-minute ES futures bars (Databento's ES.c.0 continuous contract, with columns ts_event, instrument_id, high, low, close; timestamps in UTC). Please write Python code that measures how price touches the 20 EMA on the 5-minute RTH chart. Rules:
1. Convert timestamps to US Eastern time (US/Eastern), handling daylight saving time automatically. Use RTH only, 09:30–16:15, and build 5-minute bars from the 1-minute bars, 81 a day (09:30–09:35 is bar 1). If a 5-minute slot has no trades at all (for example during the March 2020 circuit-breaker halts), that bar is empty.
2. 20 EMA = 20-period exponential moving average of 5-minute closes, alpha = 2 / 21: EMA = alpha × this bar's close + (1 − alpha) × the previous bar's EMA. RTH bars only, carried from day to day, no overnight session. Skip empty bars and leave the EMA unchanged.
3. On a day when the contract rolls (instrument_id changes), restart the EMA from that day's first close and leave the day out of the stats, because old and new contract prices don't line up; same for the first day of the data.
4. Touch: the EMA (its value after this bar closes, i.e. the value drawn on the chart) sits between the bar's high and low, i.e. low ≤ EMA ≤ high.
5. Print: the sample size, the share of days that touched the EMA, the share that touched it on bar 1, and which days never touched it; the cumulative share of days whose first touch came by bars 3, 6, 12, 18, 24 and 36; and the share of days touched in each year.
6. Away from the EMA: the whole bar is above the EMA (low > EMA) or below it (high < EMA). Consecutive bars like this form one streak, reset at each open. Print the total number of streaks, the number above and below, the average per day, the median length, the share of streaks 20 bars or shorter, the share of streaks above and below lasting at least 10, 20 and 40 bars (as a share of all streaks on the same side), and the share of days with at least one streak of 20+ bars.
Please add comments to every step explaining why it's done that way.

# 20 EMA: on the 5-minute RTH chart, how many days touch the 20 EMA? Which bar touches it first? How many bars in a row does price stay away from it?
# Input: cleaned 1-minute ES bars (Databento ES.c.0, timestamps in UTC, with an instrument_id column)
import numpy as np
import pandas as pd

df = pd.read_csv("ES-1min_clean.csv", usecols=["ts_event", "instrument_id", "high", "low", "close"])
t = pd.to_datetime(df.ts_event, utc=True).dt.tz_convert("US/Eastern").dt.tz_localize(None)
m = t.dt.hour * 60 + t.dt.minute - 570                      # minutes since 09:30
rth = df[(m >= 0) & (m < 405)].assign(date=t.dt.normalize(), bar=m // 5)   # RTH 09:30–16:15, 81 five-minute bars a day

# Build 5-minute bars as a table with one row per day and 81 columns; a 5-minute slot with no trades (the March 2020 circuit breakers) is empty
bars = rth.groupby(["date", "bar"]).agg(h=("high", "max"), l=("low", "min"), c=("close", "last"))
H, L, C = (bars[k].unstack().reindex(columns=range(81)).values for k in ["h", "l", "c"])
days = bars.index.get_level_values(0).unique()
inst = rth.groupby("date").instrument_id.first().values

# 20 EMA: alpha = 2/21, RTH bars only, carried from day to day; empty 5-minute slots are skipped
# On the first day of a new contract the EMA restarts and that day isn't counted (old and new contract prices don't line up); same for the first day of the data
ema = np.full(H.shape, np.nan)
use = np.ones(len(days), bool)
e = None
for i in range(len(days)):
    if i == 0 or inst[i] != inst[i - 1]:
        e, use[i] = None, False
    for k in range(81):
        if not np.isnan(C[i, k]):
            e = C[i, k] if e is None else 2 / 21 * C[i, k] + 19 / 21 * e
        ema[i, k] = e

touch = (L <= ema) & (ema <= H)                              # the EMA sits between the bar's high and low
first = np.where(touch.any(1), touch.argmax(1) + 1, 0)[use]  # which bar touches first; 0 if never
d = days[use]
print(f"Sample {len(first)} days, touched the 20 EMA {100 * (first > 0).mean():.2f}%, touched on bar 1 {100 * (first == 1).mean():.2f}%")
print("Days with no touch:", list(d[first == 0].strftime("%Y-%m-%d")))
for n in [3, 6, 12, 18, 24, 36]:
    print(f"Touched by bar {n}: {100 * ((first > 0) & (first <= n)).mean():.1f}%")
print((pd.Series(first > 0, index=d).groupby(d.year).mean() * 100).round(2).to_string())   # share of days touched, by year

# Away from the EMA: the whole bar above it (low > EMA) or below it (high < EMA); consecutive bars form one streak, reset at each open
streaks, longest = {"above": [], "below": []}, []
for i in np.flatnonzero(use):
    best = 0
    for side, off in [("above", L[i] > ema[i]), ("below", H[i] < ema[i])]:
        run = 0
        for x in list(off) + [False]:                       # append a False so the last streak before the close is counted too
            if x:
                run += 1
            elif run:
                streaks[side].append(run)                   # a streak ended; record its length
                best = max(best, run)
                run = 0
    longest.append(best)
up, down = np.array(streaks["above"]), np.array(streaks["below"])
both, longest = np.r_[up, down], np.array(longest)
print(f"{len(both)} streaks away from the EMA ({len(up)} above, {len(down)} below), {len(both) / use.sum():.1f} a day on average")
print(f"Median length {np.median(both):.0f} bars, 20 bars or fewer {100 * (both <= 20).mean():.1f}%")
for n in [10, 20, 40]:
    print(f"At least {n} bars in a row: above {100 * (up >= n).mean():.1f}% ({(up >= n).sum()} streaks), below {100 * (down >= n).mean():.1f}% ({(down >= n).sum()} streaks)")
print(f"Days with at least one streak of 20+ bars: {100 * (longest >= 20).mean():.1f}%")