A field is load-bearing only if a later step reads it, and Deletion Completeness fires on the diff in every mode (#326) - #333

…header nothing writes

The rung gated on the slice body containing a `### Deletes` section. Five
sites read that header — execute/SKILL.md twice, review-checklist.md three
times — and no skill in the pack ever writes it. /prd-to-issues renders
Produces, Consumes, Context, Acceptance Criteria and Assumptions, and by the
issue's Non-Goals it should not be taught to render Deletes, because that
would make reachability depend on planning-time compliance (the Context block
measures at 16 of 274 fulcrum issues).

Measured on 274 fulcrum issues: `### Deletes` appears in 0. Every one of the
five gates has been unreachable since it was written — Leveson's asynchronous
evolution, readers added and the writer never was.

The trigger is now the diff, which both steps already read:

  git diff --diff-filter=DR --name-status "$(...origin/HEAD...)"...HEAD

Verified against a scratch fixture rather than written from memory: a deleted
path prints `D<TAB><path>`, a rename prints `R100<TAB><old><TAB><new>`, and
the command exits 0 whether or not anything matched — so the prose gates on
output, not exit status. /execute carries the canonical statement and
review-checklist.md points at it; the following commit pins that the two
literals agree.

`### Deletes` survives as an optional hint naming which surfaces to sweep.

Refs #326 (Mediator change set item 2; rows 2 and 3).
Corroborated independently by
docs/solutions/architecture-decisions/deletion-unit-is-function-not-syntax-2026-08-15.md,
whose Prevention section proposed this same trigger change on a different
incident.

…st it

The trigger the previous commit installed is itself a documented command an
agent copies into someone else's repo, so it goes to the mechanism rather than
to review — CLAUDE.md § Commands a skill documents rule (b), and this suite's
own header ("Extend the suite when a skill documents another runnable command").
No new suite: #326's Non-Goals rule that out, and this shape already had a home.

Adding /execute to the census immediately found two defects in the prose the
previous commit shipped, neither visible by reading:

  1. The first draft inlined `$(git symbolic-ref … | sed …)` as the range
     endpoint — a branch NAME where only a ref is safe. That is the #298 defect
     verbatim, in prose written by someone who had just read the fix. /execute
     now carries the same detection block as the other five sites, and the same
     byte-identical **Residual:** paragraph, which it was also missing.
  2. The census's flag arm could not parse `--diff-filter=DR`, so the line
     matched nothing and dropped silently out of the endpoint check. The
     raw-`..HEAD` reconciliation is the only reason that was visible at all.

New in the suite:

  - A dedicated fixture planting one deletion, one rename, and one base-side
    ADDITION after the branch point. The main fixture deletes nothing, so it
    scores the trigger 0/0/0 — empty output there is indistinguishable from a
    broken command.
  - Three negative controls, each built by mutating the real extracted line:
    two-dot (invents a deletion for base-side work), filter dropped (the trigger
    stops discriminating and fires on any diff), and the existing bare-name one.
  - review-checklist.md is not a SKILL.md and falls outside every derived loop,
    so its copy is pinned byte-identical to /execute's canonical line. That is
    the mechanism behind review-checklist's own "same condition and same
    command" sentence.
  - The stat-triple case arm now keys on `--stat` rather than on `git diff`, and
    an unrecognised range shape fails by name instead of being scored against
    the wrong oracle.

Six mutations verified red before this landed, including reverting /execute to
the pre-#326 header gate (fatal) and dropping the filter (fatal).

Refs #326.

…half

deletion-unit-is-function-not-syntax-2026-08-15.md proposed this exact trigger
change on an unrelated incident and marked it "an /improve-pipeline candidate,
not yet filed." The previous commits filed and shipped it, so the entry now
says something false about the repo it describes.

Both halves recorded rather than one. #326's Mediator bounded the change set to
the trigger and declined the surface-list widening, so:

  - count 1 (header gate) — fixed, with the new command named
  - count 2 (surfaces are all code surfaces) — still open

The Shelf Life clause's retirement condition is a conjunction and only its first
half is met, so the entry does not retire. The clause now says so explicitly,
including the gap the wording opens: the rung fires on a diff that deletes or
renames a MODULE, which is narrower than "net-subtractive" — a prose diff that
removes a sole-site declaration without removing a file still does not trigger
it. That is this repo's own most common diff shape, and it is the strongest
candidate for a follow-up issue.

Refs #326.

…d drop the claims nothing checks

The #331 review found the new Deletion Completeness trigger reads
$BASE_REF, which Phase 1 resolves only in author-mode, so in
`/pre-merge --pr <n>` the rung could not fire: the unreachable-gate
class this branch exists to fix, one mode over. The checklist now
carries a reviewer-mode read of the same fact from the PR's file list
(GitHub's list-files endpoint, keeping `removed` and `renamed` rows),
and the same query with the upstream PR's number is how a consumed
upstream slice's deletions are read in every mode, which is the
coverage the "or an upstream slice" clause was claiming and the git
command alone did not compute.

Also from that review: the checklist cited Surgical Scope for a diff
stat that lives under Review-friendly Size (the cross-reference is
dropped rather than corrected); "recognise" in a script comment; the
anatomy bullet closed on two negations and now says what a valid
answer looks like; and the counts restated in prose and comments
("five reader sites", "0 of 274", "/prd-to-issues does not emit")
are deleted, since nothing here checks them.

… base ref, and pin the three things #333's review found unpinned

An independent probe of the reviewer-mode reach found that the
author-mode trigger, run with $BASE_REF unset, exits 0 and prints
nothing: git reads the empty left endpoint of `...HEAD` as HEAD, so
the rung's own prose reads the result as "nothing deleted, skip." The
checklist's block now opens with a `${BASE_REF:?}` guard, and the
suite runs the block both ways in the planted fixture. Writing that
guard produced its own defect first: an apostrophe inside the message
left bash looking for a closing quote, and the new assertion caught it
before it shipped.

The review's two Concerns are pinned in the same suite: the gate
sentence at both sites names the diff and not the header (reverting
it to the pre-#326 wording had left every assertion green), and the
reviewer-mode `--jq` filter, read from the checklist rather than
restated, is run against a fixture in the documented response shape.
GitHub's half of that command stays referenced to its docs.

…iew once found

Three comment lines in the new sections claimed more than the assertion
beneath them: a history of a prior review round that no commit here
carries, a pagination effect the check only establishes by presence, and
a CI runner fact nothing pins. Each now states the check it sits above.

…nce it checks, nothing more