changelog-parity gate: has_heading's early-exit reader SIGPIPEs rendered_lines under pipefail, failing correctly documented bumps

[kyle-sexton]

Defect

scripts/check-changelog-parity.sh --check-bump reports UNDOCUMENTED BUMP for a bump whose ## [<version>] entry is present, rendered, and at column one — a false FAIL in a required merge gate.

First observed on PR #2130 (markdown-format 0.11.3 -> 0.11.4, entry at line 6 of a 43 KB changelog): failed twice in CI, passed locally on Windows.

Root cause

has_heading is a pipeline under the script's set -o pipefail:

rendered_lines - | awk -v h="$heading" '
  index($0, h) == 1 { found = 1; exit }
  END { exit !found }
'

The reader exits on first match. A real changelog puts the newest heading near the top of a file larger than one stdio buffer, so rendered_lines (awk) is still writing when the reader exits, takes SIGPIPE (exit 141), and pipefail reports the pipeline — and therefore has_heading — as failed. The found heading is misread as missing.

Reproduced deterministically in an ubuntu:24.04 container at the exact CI merge commit ba4b72fb: under gawk (the runner image's /usr/bin/awk) the gate exits 1 with the exact CI error and PIPESTATUS=141 0; under mawk it passes (different write buffering wins the race), which is also why the Windows-local runs passed. The suite's 55 fixtures all fit in one buffer and can never trip it — hence PASS=55 in the same failing CI job.

Fix

Scan the whole input (drop the early exit); add a large-changelog regression fixture that fails against the early-exit reader under gawk.