refactor: two craft skills stop explaining themselves to the reader

[rookslog]

These cuts execute an in-session slop review of this suite's own prose, read
with de-slop and compressed-deliberation open, which found each of these two
files spending about half its words on its own provenance and its own limits.
No issue records the review, so the findings it acted on are listed below.

wc -w on the two skill files:

File	Before	After
skills/craft/de-slop/SKILL.md	875	748
skills/craft/compressed-deliberation/SKILL.md	650	541

The review's own targets were about 650 and about 470. The enumerated findings
below do not reach them, and nothing beyond them was cut, so both files land
short of the target by roughly a hundred words. A second pass is a separate
decision.

What the review found, and what this does about it

de-slop:

• No part of this matches a word spent 120 words arguing a repository
  decision. It keeps This skill ships no word list. and one clause of reason.
  The rest is in source/craft/de-slop.md.
• What stands behind this shape and Boundary stated the same no-standard
  proposition twice. They are one Boundary section now.
• So this skill is the generative half on its own. restated the two sentences
  above it, and is gone.
• One test settles most of them. announced that a point mattered, which is
  the shape row E-11 of this skill's own matrix names as a defect. The
  paragraph starts at Remove the sentence instead.
• How this differs from the other skills here keeps the repetition
  disagreement and the scope claims. The paragraph that duplicated Boundary
  merges into the first clause of the one that follows it, because
  Both standards needs its antecedent.

compressed-deliberation:

• What this corrects spent 211 words, seven of its bullets describing Claude
  Opus 5. The description moves to source/craft/compressed-deliberation.md.
  The skill keeps the operative sentence, which is that the rules aim at the
  whole stack rather than at the model alone.
• The evidence sentence led with what was measured and buried the instruction.
  It leads with Apply the rules ... on their reasoning and not on our evidence now.
• Conceding a correction and Surfacing a decision were headings over two
  bullets each. They are labelled bullets under The shape of a reply.

The G-row deletion is the operator-approved decision here

Moving the seven vendor bullets out of the skill deletes all seven G rows
from grounding/craft/compressed-deliberation.md. The operator approved that
directly. Three consequences, all of them mechanical once the rows go:

• The matrix drops its **Source version:** line. ADR-0026 refuses that line
  in a matrix with no G row, and ground --check enforces it by name.
• The Audited and Source text cells go empty throughout, which is the state
  grounding/craft/de-slop.md already carries.
• The **Quotation:** forbidden declaration stays, with its reason rewritten,
  because the pages it protects are still the pages behind this skill.

The provenance does not go anywhere. source/craft/compressed-deliberation.md
carries the seven statements whole, each beside the A1, A3 or A5 section
it came from, and it already held the URLs and the read dates. No ADR is added,
because no rule changed: ADR-0026 already decides the pin, ADR-0020 already
decides the declaration, and the skill still names its model build in Purpose.

Propagation

Three files named things that no longer exist, and now name what stands there:
grounding/craft/subagent-returns.md cited the seven G rows,
source/craft/proportionate-execution.md said the skill cites the vendor
pages, and the E-16 row in grounding/craft/de-slop.md still pointed at the
retired SOURCE.md. Front-matter descriptions are untouched. No editorial
stamp and no audit date was written.

Checks

npm run check exits 0. All nine: 901 tests pass, lint:docs clean,
ground --check --all clean, docs metadata clean over 37 documents, probes,
verdicts, resident, studies and editorial all clean.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 741446c8a4

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".