yq -i is not idempotent: a second run moves a comment block (tag-ruleset-canon.yml)

[hyperpolymath]

Found by the #1021 comment-preservation proof (PR #1067). Blocks Y-2 (yq for writing) in 3-practice/YAML-POLICY.adoc.

Repro — bd9313a6, yq v4.53.3

cp .github/workflows/tag-ruleset-canon.yml /tmp/t.yml
yq -i . /tmp/t.yml; cp /tmp/t.yml /tmp/p1.yml
yq -i . /tmp/t.yml; diff /tmp/p1.yml /tmp/t.yml   # non-empty

What moves

Original (lines 46–63): two comment paragraphs separated by a blank line, both before the Mint… step:

      - uses: actions/checkout@3d3c42e5… # v7.0.1

      # WHICH App's credentials belong in vars.APP_ID / secrets.APP_PRIVATE_KEY:
      # … (10 lines, the R-14 App-identity rationale)

      # An App installation token is scoped to ONE owner …
      - name: Mint an App installation token for hyperpolymath
        id: tok-user

Pass 1 keeps both before the step (comment-clean). Pass 2 moves the R-14 block inside the step, between name: and id::

oracle: jobs/converge/steps/1/name (before) → jobs/converge/steps/1/id (before) — 10 comments.

The same drift occurs under a realistic edit (yq -i pin bump). yq -o kyaml does not drift on this file (56/56 idempotent).

Candidate cause, not investigated: go-yaml v3 head/foot comment reattribution around blank lines.

Acceptance criteria

• bun prove.js .github/workflows/*.yml (tools/yaml-comment-proof) reports pass-2 drift 0 and idempotent 56/56 for the identity and pin-bump arms — by a yq version, flag, or wrapper that achieves it; or
• Y-2 stays NOT IN FORCE, with this issue cited in YAML-POLICY §4 as the reason.
• Either way the harness still passes its own controls (calibration + both mutants) in the same run.

🤖 Generated with Claude Code

https://claude.ai/code/session_01WPSJ7fBhVAMcpSffCBWUDo