實價收益 — the visible range of Taiwan's rental yield

A rental yield requires a sale price and a rent for the same dwelling. Taiwan's public transaction registry (實價登錄) records both, but for most buildings only one. This work pairs real sales with real leases in the same building across 112S4–115S2, publishes the median yield, and publishes the denominator: of 157,407 buildings with a recorded sale, only 9,843 (6.25%) can be matched at all, against a structural ceiling of 24.2% with every quality filter removed. Matchability is itself selective — roughly 1 in 9 apartment towers against 1 in 65 walk-ups — and the measurable slice is 8.2% dearer per ping and twelve years newer than the rest. Rather than assume the unobserved 94% resemble the observed 6%, the median is bounded: [0.30%, 12.00%] under no assumptions, narrowing to [1.90%, 4.27%] under assumptions stated rung by rung, including one rung that widens the interval and is published because it fails. A median's worst-case bound only leaves the support band once more than half the population is observed, so the width is a property of the registry, not of the cleaning. Three results were withdrawn after publication and are documented rather than deleted, including the project's own founding hypothesis, which is closed as unanswerable from this registry. This deposit archives the live page, the full build pipeline, the test suite and every pinned reference series, so the payload can be rebuilt and compared byte for byte. The paper is published in English (docs/HOUSING_FINAL_REPORT.pdf) and Traditional Chinese (docs/HOUSING_FINAL_REPORT_ZH.pdf). Both are generated from the same housing_data.json, and the test suite asserts that every headline figure appears in both.