\section{Protocol, eligibility criteria and coding conventions}
\label{app:protocol}

\paragraph{Pre-registered protocol.} The analysis protocol (replication file \texttt{protocol/preregistration.md}) was written on
2026-10-20 before data construction. It fixes the primary characteristics (living area, lot size, age, bedrooms, bathrooms,
garage, pool, fireplace, air conditioning, basement, storeys, distance to the central business district, distance to transit,
school quality, crime, air pollution, noise, flood risk, green space, water view or frontage, energy efficiency), the primary
effect-size family per characteristic, the models M1--M6 of Section~\ref{sec:framework}, five directional hypotheses (H1
signs; H2 $I^2$ above 75 percent everywhere; H3 no systematic difference between spatial or machine-learning marginal effects and
OLS once moderators are controlled; H4 small-study effects for signature amenities but not for area; H5 rising implicit prices of
energy efficiency and green space after 2008) and the validation design (independent double extraction of at least 15 percent of
included studies, minimum 20).

\paragraph{Eligibility.} Inclusion required an empirical study estimating a hedonic-type relationship between a residential price
measure (transaction price, listing price, appraised value or rent) and at least one identifiable property, neighbourhood,
environmental, locational, market or AI-derived characteristic, reporting at least one coefficient with a standard error,
$t$-statistic, $p$-value or confidence interval or enough statistics to recover one transparently, with a retrievable full text,
in any language. Exclusions: no residential price outcome (E1); no property-characteristic regressors or aggregate units such as
cities, tracts or housing projects (E2); theory, methods or simulation only (E3); reviews (E4, retained for the comparison of
Table~\ref{tab:competing}); no empirical regression (E5); duplicate version (E6); not the housing market, including short-term
rental nightly rates (E8); full text not retrievable (E7).

\paragraph{Conventions decided during screening and extraction.} Hedonic price-index papers estimated on micro data were
included because characteristic coefficients are usually reported; pure repeat-sales and index-methodology papers were excluded.
Chinese residential-community (xiaoqu) studies with dwelling characteristics were included and flagged; neighbourhood, district
and city aggregates were excluded. Studies reporting SHAP or partial-dependence effects from machine-learning models were kept
for the AI synthesis but their effects were not pooled with regression coefficients. Geographically weighted regression
summaries (means or medians of local coefficients) were recorded without standard errors and do not enter the pooled analysis.
Rounded-to-zero standard errors were stored as missing. Interaction terms, squared terms, standardised coefficients and
construction-year variables were flagged and excluded from pooling. Every deviation from the protocol is logged with its date
and reason in \texttt{reports/methodological\_decisions.md}.