3 년, 41 년, 한 세기 반

def span_years(first: str | None, last: str | None) -> float | None:
    """How much history stands behind a percentile, in years.

    Not decoration. These windows are wildly unequal -- HY OAS
    carries ~3 years (FRED trims the ICE BofA licence), CAPE ~146
    and rising (this docstring said 155 while the function it
    documents returned 145.7), fed funds 41 -- and a percentile
    says nothing without the ruler
    it was measured against. p12 over three years and p12 over a
    century are different claims, and a surface that prints them
    identically is asserting a commensurability it does not have.
    """
    ...


# This docstring read "CAPE 155" from 2026-08-06 to 2026-08-31 --
# (14e2586, the commit that introduced span_years itself; an
#  earlier draft of this note said 08-05, which git disagreed
#  with -- in a passage about checking claims) --
# off by a decade, in a docstring whose entire subject is stating
# your window accurately, about the very function that returns the
# right answer fourteen lines further down. The stored series
# starts 1881
# (CAPE needs ten years of earnings before its first point, so the
# ~1871 dataset yields a ~146-year window), and the live read model
# reported 145.7 the day it was corrected.
#
# Note what the fix is NOT: another hardcoded number. 145.7 would
# be wrong by next year for the same reason 155 was wrong -- the
# window GROWS. The corrected text says "~146 and rising" and
# points at window_years, which is computed. A constant restating
# a computed value is a second source of truth with no owner, and
# it always loses.


# Attached to every read-model row, next to the percentile itself:
#   row["first_date"]   = first[key]
#   row["window_years"] = span_years(first[key], row["data_date"])
#
# Two extra fields. They are the difference between publishing a
# measurement and publishing a number that looks like one.