paperlint 2.0.0

package/skills/tighten-paper/SKILL.md

@@ -0,0 +1,368 @@
+---
+name: tighten-paper
+description: "Use when a paper \"feels long / bloated / overly complex / hard to read / doesn't deliver its point / the middle drags\" — or after review rounds deposited hedge-mass. The developmental / editorial pass on a DRAFTED paper: what a real human editor thinks reading it cover to cover — the ONE point, does every section earn its place, can it be SKIMMED, is it too long for its contribution, which sections are TMI that belong in the artifact/appendix, does the conclusion pay off the intro's promise. Produces a structural verdict + a concrete cut / fold / merge / reorder plan — NOT sentence fixes. Its verdict is a required INPUT to pc-panel-review and a hard gate in harden-paper. NOT sentence craft or a writing grade (grade-paper-writing), NOT jargon stalls (the persona stall pass), NOT scientific defects (pc-panel-review / paper-adversarial-review). Run it BEFORE those on a bloated draft — no point polishing sentences in a section that should be cut."
+allowed-tools: [Read, Write, Edit, Grep, Glob, Agent, Bash(node .claude/skills/paper-pipeline/scripts/announce.mjs:*), Bash(node .claude/skills/paper-pipeline/scripts/ledger.mjs:*)]
+---
+
+<!-- vigiles:sha256:79a018d2fd2bf3e7 compiled from skills/tighten-paper/SKILL.md.spec.ts -->
+
+# tighten-paper — the developmental edit
+
+## Run me
+
+🔴 FIRST, before any other step:
+
+```
+node .claude/skills/paper-pipeline/scripts/announce.mjs tighten-paper <paper-dir>
+```
+
+An advisory pass cannot be observed failing — silence is both its error state and its normal
+state — so starting is an event, and events get written down.
+
+Sentence-level passes make a wall of text into readable text. They do **not** tell you the wall shouldn't be there
+at all. This pass is the editor who reads the whole paper once and asks: *is this the right paper, at the right
+length, delivering the right point?* Its deliverable is a **cut/fold/merge/reorder plan and a length verdict**, not
+reworded sentences.
+
+**Run it as a persona subagent too** (see `paper-pipeline/references/writing-craft.md` — the persona mechanic): an
+editor persona who reads the whole thing once, fast, the way a busy reviewer skims before deciding. A frontier model
+grading section-by-section never feels the *cumulative* drag; a persona reading cover-to-cover does.
+
+## 🔴 Run this ONCE, at the end — never after each addition
+
+The failure this rule exists for is not hypothetical; it is what produced the worst score
+`the reference paper` ever received. The body sat exactly at its 8-page limit, so every content
+addition pushed it to 9, and each time prose came out to get back — around fifty times, sixty
+characters at a go, always inside whichever section had just been edited. No single cut was wrong. The
+sum of them took the transitions, the pronoun antecedents and the honest half of the conclusion, and
+four blind readers independently scored clarity 2/5.
+
+**So: content work runs over the limit deliberately.** This pass happens when the content has settled,
+with the whole draft in view — which is the only vantage from which you can *move* a section rather
+than shave a clause. A tighten pass performed inside a single section, under pressure from a page
+counter, is not this skill; it is the damage this skill exists to undo.
+
+If you are invoked while the author is still adding material, say so and decline the pass. The right
+answer is "not yet", not a smaller cut.
+
+## 🏗 START WITH THE MECHANICAL LEG — run `structure.mjs` (added 2026-08-05)
+
+```
+node .claude/skills/tighten-paper/structure.mjs <paper.md>              # whole paper
+node .claude/skills/tighten-paper/structure.mjs <paper.md> --section=3  # one section
+```
+
+🔴 **THE SCRIPT NO LONGER FINDS ANYTHING — it is an INVENTORY (2026-08-26).** All sixteen of its
+checks moved into ESLint rules (`eslint-rules/paper-structure.mjs`). The findings you used to read
+here now come from:
+
+```
+npx eslint --no-config-lookup --config eslint.config.mjs <paper.md>
+```
+
+⚠️ Do NOT pass `--flags-only`. It is kept only because two harnesses assert on its silence, and it
+prints nothing and exits 0 whatever the paper looks like. Measured 26.08 on `the reference paper`:
+`--flags-only` printed 0 lines while ESLint on the same bytes reported **6 structural findings**
+(subsection-size ×2 · section-lead · free-section-size ×2 · block-ungraded). A mode that cannot
+speak is not a check, and reading its silence as "clean" is the defect this pipeline is named for.
+
+What the inventory prints, and why it stays a script: the outline **in order** — every heading, its
+own words, its total with subsections, its share of the body, and the `carries:`/`score:`/`verdict:`
+note the section wrote about itself, printed beside its weight. That comparison is the whole point
+and a linter cannot make it: a finding has nowhere to put a table.
+
+🔴 **Why this exists, and why word counts alone mislead.** Asked whether a paper short-changed its
+contribution, per-section word counts said no — the contribution had 38% more words than related
+work — and that was true and useless. The shape showed the defect: three of eight top-level sections
+owned **0, 28 and 38** words before their first subsection. Nothing in the toolchain asked that
+question, because every metric was per-sentence or per-document and none was per-section. The author's
+word for the result was a "salad" — a jumbled mix — and he was right while the numbers said fine.
+
+**What the script does NOT decide, and you must:** the ORDER, the NAMES, and whether a section earns
+its place. Those are judgement, and pretending otherwise would be the failure this whole pipeline
+exists to prevent.
+
+## 🔤 JUDGE THE HEADINGS AS A SET, NOT ONE AT A TIME (2026-08-05)
+
+Print every heading in order and read them as a table of contents, with nothing else. **A stranger
+should be able to reconstruct the argument from the headings alone.** Two rules:
+
+- **A heading names a thing** — a linter, a configuration, a repository, a build, a constraint —
+  never an abstraction (*answers · kinds · moves · demotion · what X removes*). Two headings that a
+  reader called "VAGUE AF" and "why not mention the linter??" survived **six** consecutive passes,
+  because every pass looked only at the section it happened to be editing. Nobody ever looked at the
+  set.
+- **A heading that names two jobs means the section has two jobs.** *"Why the prose stays, and what
+  the construction costs"* is two different questions under one number, and that is visible from the
+  heading alone before you read a word of it.
+
+## 🔀 JUDGE THE ORDER (2026-08-05)
+
+Ask what reading order a stranger actually needs, then check whether the paper delivers it. The
+specific failure to look for: **the result the venue cares about most, arriving last and smallest.**
+On `the reference paper` the workshop was about agents, all 34 of its previously accepted papers had a
+model or agent under measurement, and the paper's only agent section was 4% of the body, placed after
+two sections of caveats. Measure this — do not intuit it. If you propose a reorder, price it: say
+which cross-references break and how many.
+
+## What good papers are like (the bar)
+- **One point.** A strong paper says ONE thing and everything serves it. You can state its contribution in a single
+  sentence, and every section visibly advances that sentence.
+- **Tight beats long.** Filling the page limit is not a virtue — the best measurement/security papers are often
+  *under* it. Length signals insecurity as often as rigor. Cut everything the reader doesn't need to be *convinced*.
+- **Skimmable.** The argument comes through from title + abstract + section headings + figure captions + first
+  sentence of each paragraph alone. Structure carries the story; prose fills it in.
+- **The paper convinces; the artifact exhausts.** The paper keeps only what a reader needs to believe the claim.
+  Every per-case breakdown, every "we also checked X," every exhaustive validation sub-table → artifact/appendix.
+
+## The seven editorial questions (run in order)
+1. **The one-sentence point.** Write the paper's contribution in ONE sentence. If you can't, or it takes three —
+   that's the core problem, name it. Then check: does every section advance THAT sentence? List any that don't.
+2. **The skim test.** Read ONLY title, abstract, all headings, figure/table captions, and the first sentence of each
+   paragraph. Does the full argument come through? Where does skimming leave you lost or surprised? Those are
+   structure failures (a heading that doesn't state its point, a paragraph whose lead sentence buries the finding).
+3. **Length / verbosity audit.** Is the paper longer than it needs to be to make its point? Go section by section:
+   for each, ask *does removing this weaken the ARGUMENT?* If no → cut or fold to the artifact. Flag TMI: exhaustive
+   validation lists, defensive "we also ran X," per-case enumerations, methodology detail a reader takes on trust.
+   Give a concrete target ("§5 is ~1.5pp; it needs ~0.6pp — keep the two-way validation *claim* + one number, fold
+   the rest to the artifact").
+4. **The middle sag.** Papers drag in the middle when they ENUMERATE instead of ARGUE — a pile of parallel
+   sub-findings with no through-line. For the middle sections: is each one advancing the argument, or is it a data
+   dump the reader wades through? Recommend consolidating parallel findings and leading each with its point.
+5. **Complexity budget.** Is the paper more elaborate than its contribution warrants? A simple strong finding buried
+   under heavy machinery reads as over-engineered. Could the same point be made more simply / with fewer moving
+   parts / fewer coined terms? Name the machinery that isn't earning its keep.
+6. **Structural coherence — audit the SECTION SKELETON, not just the flow.** Section structure grows organically
+   and accretes cruft; check the top-level skeleton for these specific symptoms and propose a merged one:
+   - **Thin setup sprawl:** 3+ short sections before the first result that are all "here's the thing and how we
+     measured it" (e.g. separate *battery* + *corpus* + *method* sections). MERGE into one setup section with
+     subsections — a reader shouldn't hit four throat-clears before the payoff.
+   - **A sprawling Results section** (6–9 subsections in a flat list) → REGROUP into ~3 subsections **named for the
+     argument's steps** (e.g. the spelling→grammar→effect ladder), not for the experiments.
+   - **A "What actually works" / "Discussion of implications" section that duplicates** the results + discussion —
+     usually added because the point wasn't landing elsewhere. FOLD it into the results' close or Discussion.
+   - **Redundancy:** any point/stat/definition stated in 2+ sections — full once (first use), one-clause
+     back-reference after.
+   Output the CURRENT skeleton and a proposed CLEAN one (fewer, better-organized sections). A typical clean shape:
+   Intro · Background/threat · **one merged setup** · Results (subsections = argument steps) · Discussion ·
+   Related work · Ethics · Conclusion.
+   **But fewer sections ≠ walls of text.** Lean means *organized*, not monolithic: every merged/long section KEEPS
+   subsections and bold run-in heads — they're the wall-breakers (see writing-craft.md's wall-check). Don't trade a
+   choppy 11-section paper for an 8-section paper of unscannable blocks.
+7. **The payoff check.** Does the conclusion deliver the exact thing the intro promised — no more (no new claims), no
+   less (no anticlimax)? Do intro and conclusion agree on what the paper is about?
+
+## Aggression — the default verdict for a sentence is DELETE
+The failure this skill keeps hitting is **timidity**: it concludes "the skeleton is clean, nothing major to cut"
+and trims lightly, so the paper stays at the page limit and still *feels* fat (GateBench 2026-07-25: the author
+called it "fat for no reason" after a pass that cut almost nothing and left it 10pp). Fix the posture:
+- **Clean skeleton ≠ tight paper.** A perfectly-organized section list can still be 30–40% fat prose. "No whole
+  section to cut" is **NOT a stopping condition** — go INSIDE each section and cut at the paragraph and sentence
+  level. A tightening pass whose conclusion is "structure is already clean" has not done its job.
+- **Burden of proof is on the words, not the editor.** Do not ask "can I justify removing this?" Ask "does this
+  sentence change what the reader BELIEVES or can DO?" If not — cut. The default verdict for any sentence is
+  DELETE unless it visibly advances the one-point.
+- **Each section justifies its LENGTH, not just its existence.** Existence-justified but 2× too long is still a
+  fail. Budget each section from its job (threat model ~0.5pp not 1.5; related-work ~0.75pp; a discussion that
+  re-states results ~0). Over budget → cut to budget.
+- **The fat inventory (hunt these in every section):** background the reader already knows; "we also ran/checked
+  X" reassurance; per-case enumerations (→ artifact); hedge-stacks (→ one scope sentence); a sentence that
+  restates the previous one in new words; throat-clearing lead-ins ("It is worth noting that…"); any paragraph
+  whose deletion the ARGUMENT would not notice.
+- **Measured, not vibes.** Report before→after LINES (or words) per section and total. **A tightening pass that
+  does not move the line count materially DID NOT HAPPEN** — say so and cut harder. On a paper that "feels fat,"
+  −20–30% of body prose is a normal result, not an extreme one.
+- **"It's all load-bearing numbers" is the SECOND timid excuse — reject it too** (the first is "the skeleton is
+  clean"). A paragraph can be FULL of protected numbers/cites and STILL be over-explained. Load-bearing protects
+  the CLAIM + its number — NOT the three sentences that motivate, teach, and restate it around them. Method:
+  extract the claim+number into ONE tight sentence, cut the surrounding explanation. **Dense ≠ uncuttable** — a
+  measurement paper's fat HIDES as explanation of real numbers, which is exactly why "it's all numbers" lets it
+  survive a lazy pass (GateBench 2026-07-25: a pass cut only 5% hiding behind this; the author correctly pushed
+  back that whole sections were over-explained, not just number-dense).
+- **"Fold to the artifact" is the PRIMARY length lever for a measurement/benchmark paper — not a last resort.**
+  FIRST verify the artifact exists and recomputes the numbers (it usually does for these papers). Then per-case
+  breakdowns, secondary analyses, exhaustive robustness sub-results, ablation walkthroughs, and every "we also
+  checked X" MOVE there: keep the headline number + one pointer ("the artifact's per-guard matrix reports the
+  rest"), cut the in-paper walkthrough. This is NOT dropping a claim by omission — the public artifact recomputes
+  it. **A dense benchmark paper that folds NOTHING to its own artifact has not been tightened.** Make a fold list
+  first, before any sentence-level cutting.
+- **Length is calibrated to CONTRIBUTION WEIGHT, not to the page limit.** A moderate finding + a moderate-complexity
+  benchmark does not need to sit at the limit; being over-length for the weight of the contribution is itself a
+  defect — it reads as padding even when each sentence is individually "true." Right-size to the contribution's
+  natural length, then stop. Ask: "for how big this finding actually is, how many pages does it deserve?" — and cut
+  to THAT, not to the ceiling.
+- **Apply it — don't defer it (see caveats for why this is safe).** When invoked to TIGHTEN (not merely audit for
+  harden/pc-panel input), APPLY the prose/TMI cuts directly and report the measured delta. A plan with no applied
+  cuts is only acceptable for a pure pre-review audit; "tighten this, it's fat" means CUT.
+
+## 🔴 Write the per-section verdict INTO the paper, not just into chat
+
+A verdict that lives in a chat log is gone by the next session, and the next session then re-derives
+it or ignores it. **Every section and subsection gets a one-line verdict comment immediately above its
+heading**, in the paper source, stamped with the date and the skill that produced it:
+
+```
+<!-- TIGHTEN 2026-08-06 · score 8/10 · verdict: KEEP · 190w
+     carries: tools get close but none resolves a NAMED rule against a committed config
+     because: it is the only place the tool-ecosystem gap is closed; a reviewer who skips it
+              asks "hasn't this been done" and gets no answer -->
+<!-- TIGHTEN 2026-08-06 · score 5/10 · verdict: SHORTEN -89w · 210w
+     carries: why a linter for the rules file cannot work
+     because: the point lands in two sentences; the rest re-establishes what §2 already did -->
+<!-- TIGHTEN 2026-08-06 · score 2/10 · verdict: MOVE→artifact · 45w
+     carries: nothing the reader must hold to judge the contribution -->
+```
+
+🔴 **Before assigning any `MOVE→` verdict, read `paper-pipeline/references/body-vs-appendix.md`.**
+It holds the one rule that decides where a block lives — quoted from the ACL/ARR guidelines, not
+invented — plus the ordered list of what to cut first when the body overruns, and the asymmetric
+case that matters most: **a qualification narrowing a number in the body belongs in the body**,
+however tight the page count. Moving it out does not fix the overclaim, it hides it.
+
+**Three fields, all three required, on every heading — and in an appendix, on every bolded block.**
+
+| field | what it means |
+|---|---|
+| `score: N/10` | how much **a reviewer's decision to accept** depends on this. NOT how interesting it is, not how hard it was, not how much you like it |
+| `verdict:` | one of **KEEP · SHORTEN −Nw · MOVE→§x/appendix/artifact · MERGE→§x · CUT** |
+| `carries:` | the belief it holds, in one clause. If you cannot write it, that IS the finding and the verdict is not KEEP |
+
+🔴 **Why a score, when there was already a justification.** A free-text note is a section arguing its
+own case, and a section always wins that argument: there is no scale to rank it against its
+neighbours, and no slot in which the answer can come out negative. Over one paper, 23 such notes were
+written and **not one ever concluded "cut me"**. Author, 2026-08-06, reading the built PDF:
+*"reading the paper gives the impression — hmm, is this section really needed"* — the notes had asserted
+value on every page and compared nothing.
+
+The pairing is the mechanism: **a low score beside a KEEP verdict is a visible contradiction**, and
+`structure.mjs` reports it. Prose cannot contradict itself; a number can.
+
+**Calibration, so the scale does not drift to "everything is an 8":**
+- **9–10** — remove it and a reviewer cannot assess the contribution. The claim, its evidence, its price.
+- **7–8** — removing it costs a reviewer's confidence: the honest bound, the nearest competitor, the cost.
+- **5–6** — a reader who already believes you wants it. Full protocols, per-item tables. This is an
+  **appendix** score, not a body score.
+- **3–4** — real work, nobody asks for it. Its home is the artifact.
+- **0–2** — duplicates something reachable, or was written for an argument the paper no longer makes.
+
+**Anything scoring below 5 that sits in the BODY is a page-limit bug**, and the verdict says where it goes.
+
+🔴 **The scale does not apply to integrity content, and a low score there is never a reason to cut.**
+Attribution, ethics and disclosure statements, licence and availability notes, correction records,
+and the admission of a limitation all score low by construction — no reviewer's accept/reject
+depends on them. That is what they are *for*. Score them honestly and **write `verdict: KEEP`
+anyway**, with the reason.
+
+Observed 2026-08-06, first run of this format: a block crediting a section title's metaphor to a
+named Supreme Court dissent scored 2/10 and came back `CUT`, described as "a courtesy footnote, not
+evidence". Deleting the credit while keeping the borrowing is not a saving. The pass was applying
+the scale exactly as written, which made it a defect in the scale.
+
+🔴 **A `MOVE` or `CUT` verdict on an appendix block is INVALID unless the note says what points at
+the block.** State the pointer, or state that a search found none — and if one exists, the verdict
+is KEEP until the pointer is rewritten first. This is not a nicety: an executed MOVE leaves the body
+saying "see Appendix B" and Appendix B saying nothing, which reads as thorough and sends the reader
+into a hole. Worse than the words it saved.
+
+Measured, first run of the format: the pass proposed **686 words of cuts across nine blocks, and
+every one had to be reversed.** Four were pointed at from the body or Limitations. One held the only
+citation of a reference — moving it would have left an uncited entry in the bibliography, which the
+pass itself noted and proposed anyway. One was the working behind a bound that had been promoted
+into the abstract that morning, so "unreached" was true when the pass started and false when it
+finished. One was a population reconciliation, in a paper whose single loudest complaint is that its
+populations are never related. Two were the block that *points at the artifact*, and an attribution.
+
+**Nine for nine is not bad luck, it is a missing input.** A pass that grades blocks one at a time
+cannot see what points at them, so it will keep proposing exactly these cuts. Hand it the
+reachability map, or require the pointer check in the note — the second is cheaper and it is what
+this rule does.
+
+**Same day, two more verdicts had to be overruled**, both `SHORTEN` on subsections a word threshold
+had flagged, both scored 7/10 and 9/10, and in both the verdict was derived *from the threshold*
+rather than from reading the text — one of them on a passage carrying a standing note that says the
+fix there is expansion, not compression. **A size finding says LOOK HERE. It does not say what to
+do.** A pass returning SHORTEN on high-scoring sections it has not re-read is doing arithmetic, and
+its verdicts should be treated as unread.
+
+Rules for these comments:
+- **Every section AND every appendix block.** Demanding one note per heading asks nothing of an
+  appendix, which is one heading with thirty-five arguments under it — and that is exactly where
+  unpriced prose accumulates, because nothing there costs a page.
+- **Stale beats absent, but not by much.** A comment whose date precedes the last substantive edit to
+  its section is stale; re-grade it rather than trusting it.
+- **A verdict nobody executed is a note, not a decision.** `structure.mjs` reports blocks still
+  carrying a non-KEEP verdict — a paper full of unexecuted CUTs is worse than one never graded,
+  because it reads as decided.
+- They are working notes: stripped before typesetting, like every other `<!-- -->` block.
+- The chat summary still gets written — but it is the *derivative*, and the paper is the record.
+
+## 🧱 Enforce subsections — a section without them is a wall
+
+**A section that runs past roughly 700 words or five paragraphs with no `###` under it is a wall**,
+and a wall is where TMI hides: nothing announces what each part is for, so nothing can be judged
+individually and nothing gets cut. Readers skim headings; if a section offers none, its middle is
+invisible.
+
+For every such section, either propose the subsection split (one heading per belief the reader carries
+out, named by the belief and not by the topic) or explain why the section is genuinely one indivisible
+argument. Introductions and conclusions are the usual legitimate exceptions; a results or related-work
+section almost never is.
+
+Mechanical leg: `paper-lint`'s `checkPaperWalls()` reports oversized unsubsectioned sections on the Stop
+pass, so this does not depend on someone remembering to run the skill.
+
+## Output — the verdict AND the applied cuts
+Produce:
+- **The one-sentence point** (as you could best extract it) + whether the paper delivers it clearly (yes / buried / no).
+- **Length verdict:** is it too long? by roughly how much? Is it at the page limit because it's full of argument, or
+  because it's padded?
+- **The cut/fold/merge/reorder plan:** a ranked list — for each: section, action (CUT / FOLD-TO-ARTIFACT / MERGE
+  with X / REORDER / TIGHTEN), why, and rough length saved. Lead with the highest-impact.
+- **The middle-sag diagnosis** and the fix (consolidate these findings, lead each with its point).
+- **Skim-test failures** (headings/lead-sentences that don't carry the argument).
+
+## Record the verdict
+
+🔴 LAST step, once the deliverable exists:
+
+```
+node .claude/skills/paper-pipeline/scripts/ledger.mjs record tighten-paper <paper-dir> FINDING <count> <report-path>
+node .claude/skills/paper-pipeline/scripts/ledger.mjs record tighten-paper <paper-dir> ABSTAINED <reason> "<one line>"
+```
+
+**FINDING** is the normal row — `<count>` is the number of CUT / FOLD / MERGE / REORDER actions in
+the plan and `<report-path>` is the plan itself. Add `--blocking` when the paper is over its limit
+and the plan cannot get it under.
+**ABSTAINED** — `no-witness`: a cover-to-cover read produced no action at all, which on a real draft
+should be rare. `blocked`: the declined pass of the section above — content is still being added, so
+there was nothing to read whole.
+
+🔴 **This is the check with the documented history of never saying no**: a tightening pass once
+returned KEEP on 80 of 81 sections and still counted as run. It is also one of the three failures
+that deleted `PASS` — the word made "I read it all and nothing moved" indistinguishable from "I did
+not really read it". `status.mjs` reports a check that has never recorded a finding, on the
+mutation-testing principle that a test which kills no mutant is not a test. So an `ABSTAINED
+no-witness` here owes a note saying what was read and why nothing moved.
+
+The mechanical half of this skill files separately, under `tighten-paper/structure` — one check, one
+row. Do not record a `structure.mjs` result here.
+
+## Compose
+Run this FIRST on a draft that "feels bloated" — before grade-paper-writing / the stall pass / pc-panel, because
+polishing sentences in a section that should be folded to the artifact is wasted work. After the plan is applied,
+THEN run the sentence-level passes on what remains. On a page-limited paper, this pass is also what *buys* the space
+a sentence-level readability pass needs (cut a TMI section → the core mechanism gets room to breathe).
+
+## Honesty / scope caveats
+- **Separate the two risk classes — they are NOT the same, and conflating them is what makes this pass timid.**
+  (1) **Prose/TMI cutting** — deleting fat sentences, folding enumerations to the artifact, collapsing hedge-stacks,
+  re-voicing academic sentences plainly — removes words that don't change what the reader believes. **LOW risk;
+  apply it aggressively even the day of the deadline.** Timidity here IS the failure, not the safe choice.
+  (2) **Skeleton surgery** — merging, reordering, or deleting whole sections — can destabilize a done paper. **THIS
+  is the deadline judgment call:** propose it, get author sign-off, hold for camera-ready if unsure. Never let
+  caution about (2) suppress (1) — a "hold big cuts for later" reflex applied to *prose* fat is exactly how a
+  "feels fat" paper survives a tightening pass unchanged. Don't silently do (2); do (1) hard.
+- "Fold to the artifact" only works if the artifact actually exists and can hold it — verify (ties to the
+  availability check in `harden-paper`).
+- Cutting changes claims by omission. Anything cut/folded still gets the claim-preservation diff (the `claims` row) — a
+  removed hedge or caveat is a claim change.