docs(arch-v2): record operator verdicts on S-H DECISION-NEEDED #1 and #2

[artyhoo]

Summary

Records two operator verdicts (2026-08-07) on S-H's open forks, where the consumer reads them —
not by editing the S-H research patches, which are append-only and read-only for later sessions per
the Artifact Ownership Contract.

fork	verdict	recorded at
DECISION-NEEDED #2 — the [W] denominator	Option A — [H] operative from S-H onward, [W] historical	spec's binding denominator convention + the N1 falsifier cell
DECISION-NEEDED #1 — corpus drift	Option B — host-side snapshot of the per-turn billing projection before any prune	S-D′ kickoff, as an INPUT CONDITION

DECISION-NEEDED #3 was answered by the /context paste and shipped in #1249. #4, #5, #6
remain open by design and are untouched here.

Recording these does not unblock S-D′: that stage stays not dispatchable for two independent
reasons — its own §5 fork (resolved by the operator, but recorded only in the follow-up rev-6 PR)
and the two-gate S-E + S-H merged dependency, of which S-E has not merged.

Changes

docs/superpowers/specs/2026-08-06-pipeline-token-economy-design.md — two annotations:

1. After the binding denominator convention: [W] is HISTORICAL from S-H onward. Stated as a
   constatation, not a preference — [W]'s 169-session subset is physically unrecoverable,
   because worktree pruning deleted 34 project directories together with their transcripts, so «the
   169-session subset cannot be re-selected by re-running anything»
   (…-s-h-turn-attribution-p3d-p11.md:112, cause :92). Three binding consequences spelled out:
   existing [W]-tagged shares stay readable as history but are not comparable past 2026-08-07;
   no new figure may be tagged [W]; every [W]-defined threshold is re-adjudicated on [H] or
   marked un-adjudicable. Non-convertibility is restated, citing the patch's own caveat at :168.
2. The one [W]-defined threshold — N1's falsifier — annotated in place: struck through, marked
   un-adjudicable as written, given its operative [H] restatement, and carrying the standing
   reading 16.5% > 5% → the discipline text STAYS (:154, Option-A consequence :162).

.claude/orchestrator-prompts/arch-v2-context-pipeline-s-d-prime/kickoff.md — an INPUT
CONDITION in the header region, beside the existing two-gate dependency:

• Why: this stage's before/after measurement is invalid on a corpus that shrinks under it —
  −23.5% between the seed and S-H (247 → 189 files, 99 → 65 project dirs), a population change, not
  drift (…-s-h-turn-attribution-p3d-p11.md:73-95).
• Binding: the billing-projection snapshot is captured before any worktree prune, and this
  stage's baseline cites the snapshot it used.
• Form — reuse, not build: copy the shape of the existing origin/data/metrics channel (one
  appended CSV row per day on a data branch, outside main history) but not its mechanism — that
  is a CI cron (.github/workflows/metrics-collect.yml) collecting GitHub popularity counters
  (date,stars,forks,clones_…), and the corpus lives at ~/.claude/projects, which CI cannot
  reach. The snapshotter is host-side by construction, not by preference.
• Scope fence: authoring the snapshotter is explicitly not this stage's deliverable — the
  clause states the precondition and its form; the owning stage is named at dispatch.
• Footnote — one consumer, not two: with S-D′ §5 resolved to Option A (operator verdict
  same date; the resolution itself lands in kickoff rev 6, a separate PR), the ADR-8 A/B arm is
  descoped, so the surviving consumer of a stable baseline is this stage's own before/after.
  ADR-8's 20-dispatch window travels to the follow-on stage, which inherits the same condition.

Prior-art consult

• The commit carries a Prior-art: trailer — escape-hatch form. Not a capability commit under
  the CLAUDE.md mechanical definition: no new dependency, no file under packages/.
• Build-vs-reuse was actually exercised, not asserted: the own-stack channel
  (origin/data/metrics + .github/workflows/metrics-collect.yml) was found and read before
  any mechanism was proposed, and rejected on a stated, checkable ground (CI has no path to
  ~/.claude/projects) — the #own-stack-blind-spot counter from
  build-first-reuse-default.md §1.1. Its form is
  reused; only the mechanism is new, and this PR does not build it.
• No new capability area → no new SSOT entry; no existing entry matched → no Last reviewed
  touch.

Test plan

• npx vitest run packages/core/principles/ — 36 files / 348 passed, 1 skipped.
• bash .claude/hooks/check-kickoff-traps.sh …/arch-v2-context-pipeline-s-d-prime/kickoff.md →
  EXIT 0 (§4 T-enumeration intact; this PR does not touch §4).
• markdownlint — 0 errors on both files.
• 600-line markdown gate — spec 564, kickoff 322.
• Pre-commit + pre-push hooks ran clean (no --no-verify).
• Every citation introduced by this diff was verified mechanically against the cited line,
  not from recall — a sed -n '<n>p' | grep probe per citation. One citation failed the probe
  and was tightened before commit: :111-112 → :112, because the quoted clause lives entirely
  on :112 and :111 only carries the sentence's lead-in.

Provenance

n/a — not a stage PR: this executes no kickoff and produces no stage deliverable; it records operator decisions into an existing spec and an existing kickoff.

Review findings

Self-review only, and stated as such. The substance is operator-authored (both verdicts and their
grounds arrived in the handoff); my contribution is placement, wording and citation. The failure
mode this workstream has actually exhibited — nine REVISE rounds on the S-H addendum — is derived
quantities across mismatched populations, and this diff derives nothing: it introduces no
share, ratio, product or magnitude word of its own. Every figure it names (−23.5%, 247 → 189,
99 → 65, 34 dirs, 16.5%, 5%, 35.2%, 189 + 722) is quoted from a merged patch with its line, and the
mechanical probe above is the check that they point where they claim.

Residual risk, named honestly: the ~~strikethrough~~ on the N1 falsifier cell is a presentation
choice, not a semantic one. A reader skimming the table sees the cell struck and might stop before
the operative [H] restatement that follows in the same cell. The alternative — deleting the [W] text
outright — loses the record of what the falsifier used to say, which the verdict explicitly wants
preserved as history. Flagged rather than silently resolved.

Fidelity verdict

FIDELITY: skipped — not a stage PR: no kickoff is being executed and no stage deliverable is produced; this records two operator verdicts into an existing spec and kickoff, with every citation verified mechanically (see Test plan).

§1.7 Forward-check applied

• .claude/rules/build-first-reuse-default.md:66 — the own-stack channel was surveyed
  (origin/data/metrics, .github/workflows/metrics-collect.yml) before a mechanism was named,
  and the verdict is REUSE-the-form / reject-the-mechanism on a checkable ground, not «ours is
  special». context7 correctly not consulted — its tooling caveat scopes it to library API docs.
• .claude/rules/kickoff-staging-placement.md:28 — the S-D′ kickoff edit follows the binding
  sequence: author → merge to staging → only then dispatch. Nothing is dispatched by this PR,
  and S-D′ is independently non-dispatchable, so #dispatch-before-staging cannot fire here.
• .claude/rules/ai-laziness-traps.md:56 (T3), :159 (T20) — no prose-only claim: every
  citation carries the line it came from, and the mechanical probe is quoted in the Test plan rather
  than asserted.
• .claude/rules/doc-authority-hierarchy.md:15 — the spec keeps its own authority; this PR adds
  scoped annotations inside it and does not re-declare what the spec is authoritative for. The
  kickoff stays header-exempt per §2's tracked-but-exempt carve-out for
  .claude/orchestrator-prompts/*/kickoff.md.
• .claude/rules/no-paid-llm-in-ci.md:20 — nothing added to CI.

§1.7 Backward-check applied

Class of this change = a denominator-tag verdict that invalidates figures already published under
the superseded tag.

Disclosure — the first draft of this sweep was wrong, and a probe caught it. I initially
enumerated from recall: it named the S-E kickoff and both P14 patches as swept surfaces and omitted
the S-H kickoff. Running the actual command changed three of six rows. The corrected sweep below is
the command's output, not a recollection. This is the #backward-check-restates-not-sweeps failure
mode (ai-laziness-traps.md T21) caught before ship rather than
after.

Population — every tracked file carrying the tag, grep -rln '\[W\]' docs/ .claude/ → exactly
four:

1. docs/superpowers/specs/2026-08-06-pipeline-token-economy-design.md — SWEPT (this diff): the
   convention annotated, and its single [W]-defined threshold (N1) annotated in place. Its other
   occurrences are scope tags on stage rows (P1, P3, P5, P6, P7, P11, P13, P14),
   saying which axis a stage acts on — not measured shares, so no re-tagging is owed; the blanket
   «no new figure may be tagged [W]» governs what comes after.
2. docs/meta-factory/research-patches/2026-08-07-s-h-turn-attribution-p3d-p11.md — NOT SWEPT, by
   ownership: it is the patch that raised the fork, and research-patches are append-only +
   read-only for later sessions. Already internally correct under the verdict — it tags every share
   [H] at the point of use and prints the run-time warning.
3. .claude/orchestrator-prompts/arch-v2-context-pipeline-s-h/kickoff.md:39 — GAP-FOUND, and
   deliberately left: this is the line «sized against WRITE [W]», i.e. the very instruction
   fork feat: self-application + meta-factory foundation (Phase 0.5–3) #2 reports as unsatisfiable. S-H is closed (arch-v2: stage S-H — host-side measurements (P3d per-turn attribution SSOT + P11 Explore/Plan probe + P14 harness price list) #1239 + docs(s-h): P14 /context addendum — the operator paste that answers DECISION-NEEDED #3 #1249 merged); its kickoff is the
   historical record of what was asked, and rewriting it would erase the evidence that the stage
   sized on [H] and said so. The correction lives where it belongs — in the patch the stage
   produced, and now in the spec.
4. docs/meta-factory/research-patches/2026-08-07-s-h-p14-context-addendum.md:306 — SWEPT-CLEAN:
   its one hit names the [W]/[H]/[D] convention as a GAP-FOUND surface inside its own §1.7. A
   reference to the convention, not a share tagged with it.

Three surfaces I had expected to be in the population are not, verified rather than assumed:
…-s-h-harness-remainder-p14.md (grep -c → 0), the S-E kickoff (0), and the S-D′ kickoff (0) —
the last is swept by this diff on the sibling basis (corpus population), not on the tag.

Adjacent, not in this class: docs/superpowers/specs/2026-07-31-arch-v2-context-pipeline-design.md
ADR-3 carries no [W]; its 29-39% band is a denominator-choice question, tracked as S-H
DECISION-NEEDED #6 and routed by the operator's package into the #4 recalculation stage.
Untouched here deliberately — folding it in would settle #6 silently.

One surface (the S-H kickoff) is left with a known, named gap and a stated reason. No surface is
left silently inconsistent.