docs(arch-v2): S-D′ kickoff rev 6 — §5 resolved to Option A, ADR-8's A/B arm re-homed to S-K

[artyhoo]

Summary

Records the operator's resolution of the S-D′ kickoff's §5 PARKED DECISION-NEEDED (verdict
2026-08-07): Option A — the ADR-8 A/B arm is descoped from S-D′. Follow-up to #1250, which
recorded the sibling verdicts on S-H forks #1 and #2.

The stage's first cold Phase -1 returned STOP and parked this fork; rev 6 lifts that blocker.
S-D′ still cannot start — but now for exactly one reason instead of two: the two-gate
S-E + S-H merged dependency, of which S-H is merged (#1239 + #1249) and S-E is not.

Changes

.claude/orchestrator-prompts/arch-v2-context-pipeline-s-d-prime/kickoff.md → rev 6

• Header rev marker — records what moved and that the §5 blocker is lifted.
• §1 item 3: PARKED → DESCOPED. The four Phase -1 findings stay verbatim as the record of
  why; the executor instructions invert from «held with it» to «their absence is conformant,
  not a shortfall». The stage now ships deliverables 1, 2 and 4.
• §4 T2 re-scoped, not deleted. ADR-8's A/B baseline leaves with the arm, but S-D′ keeps its
  own before/after baseline, which stays subject to «captured, not described» — and now cites
  the pre-prune billing-projection snapshot per the INPUT CONDITION landed in docs(arch-v2): record operator verdicts on S-H DECISION-NEEDED #1 and #2 #1250. Deleting T2
  outright would have dropped a live obligation along with the descoped one.
• §5 flipped to RESOLVED. The fork text is kept verbatim — a resolution is only auditable
  next to the alternative it declined.
• §5 INCONCLUSIVE — CLOSED, with evidence (below).
• §6 (new): the S-K stub.

.claude/orchestrator-prompts/arch-v2-context-pipeline/kickoff.md — the stub made real rather
than promised: an S-K row in the stage table, the S-D′ row updated (Implements drops ADR-8 →
ADR-1), and the «ADR-8 is NOT orphaned» paragraph extended with the second recorded deviation
and the new host. A stage that exists only inside another stage's parked section is not
discoverable by /pipeline; this is what makes the descope operational rather than narrative.

The INCONCLUSIVE, closed

The open question was whether aif creates the task id before or after the dispatch prompt is
composed. Option B needed «before»; the answer is after:

• the kickoff ships as the description field inside the POST /tasks request body —
  packages/runtime-bridge/src/AifHandoffBackend.ts:231-239;
• the id exists only in that call's response — :249,
  const taskId = (createResult as { id: string }).id;
• the only post-create mutation is PUT /tasks/{id} {paused:false} (:260); the sole other write
  is a best-effort DELETE rollback on failure (:263). No request edits description after
  creation — confirmed by enumerating every _rest( call in the file (:132, :231, :260,
  :263, plus the private helper at :336), not by reading the happy path.

So a parity-selected arm cannot live in the dispatch prompt without a two-phase dispatch redesign —
independent grounds for Option A, beyond the four findings already recorded.

Caveat carried into the artefact, because it bounds the claim: verified client-side only.
packages/api was not audited for whether a paused task's description is editable after
creation. If it is, a two-phase dispatch gets cheaper — that is an entry question for S-K, and it
does not reopen S-D′.

Discrepancy found while executing the instruction — recorded, not silently absorbed

Option A's own text instructs that §3's «a PR that ships maps with no executable A/B arm fails the
stage's purpose» must be rewritten out. That sentence does not exist — not in §3, not
anywhere in the file.

Established by exhaustive grep rather than by reading §3:
grep -nE 'A/B|ADR-8|experiment|ledger|arm|purpose|evaluation' over the whole kickoff returns, in
the entire §3 region (lines 165-264 at rev 5), only the #optout-as-reflex note about
install-fingerprint snapshots — nothing tying the stage's purpose to an executable arm. The A/B
framing lived in §1 item 3 and §4 T2, and those are what rev 6 edits.

Option A's consequence is therefore satisfied by construction. The discrepancy is annotated inline
in §5 so the next reader does not re-hunt for a phantom line — an instruction pointing at absent
text is a defect in the fork's own statement, and the cheapest place to catch it is the file it
misdescribes.

Prior-art consult

• Prior-art: trailer present — escape-hatch form. Not a capability commit: no dependency, no
  file under packages/.
• No new capability area → no SSOT entry; no existing entry matched → no Last reviewed touch.
• context7 N/A per build-first-reuse-default.md §3
  tooling caveat — library API docs, not this problem class.

Test plan

• npx vitest run packages/core/principles/ — 36 files / 348 passed, 1 skipped.
• bash .claude/hooks/check-kickoff-traps.sh on both edited kickoffs → EXIT 0 each
  (§4 T-enumeration intact; the S-D′ §4 edit preserves the Active traps line and both
  T-SDP-* definitions).
• markdownlint — 0 errors on both files.
• 600-line markdown gate — S-D′ 384, umbrella 536.
• Pre-commit + pre-push ran clean (no --no-verify).
• The AifHandoffBackend.ts line citations verified against the file at origin/staging, and
  the «only post-create mutation» claim verified by enumerating every _rest( call rather
  than sampling.

Provenance

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

Review findings

Self-review only, stated as such. The verdict and its grounds are operator-authored; this PR is
placement, wording and one verification.

Two judgment calls worth flagging rather than burying:

1. §4 T2 was re-scoped, not removed. The literal reading of «descope the A/B» would delete the
   T2 clause naming the ADR-8 baseline. That would also delete the «captured, not described»
   obligation on S-D′'s own before/after, which survives Option A and is precisely what docs(arch-v2): record operator verdicts on S-H DECISION-NEEDED #1 and #2 #1250's
   INPUT CONDITION exists to feed. Splitting the clause is the reading that keeps both facts true.
2. The umbrella kickoff was edited — a second file, beyond the literal «S-D′ kickoff rev-6».
   Justification: the operator's package names «add the follow-on stage stub» as part of this PR,
   and a stage stubbed only inside a parked section of another stage's kickoff is invisible to
   /pipeline. Flagged because it is a scope judgment, not an instruction I was given verbatim —
   CLAUDE.md PR strategy is the rule I am reading myself against here.

Residual uncertainty, one item: the stage letter S-K. Letters A-I are taken (incl. S-D′), and I
picked the next free one rather than S-J to avoid a J/I misread in dispatch strings. That is a
naming preference with no evidence behind it — rename freely if the umbrella has a convention I did
not find.

Fidelity verdict

FIDELITY: skipped — not a stage PR: no kickoff is being executed and no stage deliverable is produced; this records one operator verdict into two existing kickoffs, with the one factual claim (task id postdates the dispatch prompt) verified against source and quoted with file:line.

§1.7 Forward-check applied

• .claude/rules/kickoff-staging-placement.md:28 — the binding sequence is author → merge to
  staging → only then dispatch. This PR dispatches nothing, and S-D′ remains gated on S-E, so
  #dispatch-before-staging cannot fire. The rev-6 text must reach staging before anyone acts on
  the descope — which is exactly what merging this PR does.
• .claude/rules/attention-is-not-a-mechanism.md:15 — the descope is not «the executor will
  notice the arm is gone»: §1 item 3 states the inverted instruction explicitly, §4 T2 states which
  half survives, and §6 names the destination. Each is a written instruction at the point of use,
  not a reader's inference.
• .claude/rules/ai-laziness-traps.md:56 (T3), :159 (T20) — the one new factual claim carries
  its file:line and was verified by enumerating every _rest( call, not by reading the happy path.
  The phantom-§3-line discrepancy is reported with the command that established it.
• .claude/rules/no-paid-llm-in-ci.md:20 — nothing added to CI.
• .claude/rules/doc-authority-hierarchy.md:15 — kickoffs stay header-exempt per §2's
  tracked-but-exempt carve-out; no authority claim is added or moved.

§1.7 Backward-check applied

Class of this change = a descope that moves a deliverable off a stage, where the moved thing is
referenced from other surfaces.

Disclosure — my drafted sweep was wrong, and the probe caught it. Again. The draft enumerated
five surfaces from recall and missed four real ones; running
grep -rln 'ADR-8' docs/ .claude/ install.sh returned nine files. This is the second time in
two PRs that the mechanical enumeration overturned a recalled one
(ai-laziness-traps.md T21 —
#backward-check-restates-not-sweeps). The sweep below is the command's output.

Swept by this diff (2):

• .claude/orchestrator-prompts/arch-v2-context-pipeline-s-d-prime/kickoff.md — header, §1 item 3,
  §4 T2, §5, new §6: all five sites that carried the arm.
• .claude/orchestrator-prompts/arch-v2-context-pipeline/kickoff.md — the S-D′ row's Implements
  column (ADR-8 → ADR-1), the new S-K row, and the «ADR-8 is NOT orphaned» paragraph, which would
  otherwise still assert S-D′ as the host. Not the file the instruction named — the surface a
  restatement-shaped check would have missed.

GAP-FOUND, routed not edited (2):

• docs/superpowers/specs/2026-08-06-pipeline-token-economy-design.md — §1.5 (:111-122) and the
  P13 row (:363) still describe S-D′ as running under ADR-8's protocol with the rev-4
  deviation. A second recorded deviation is a spec-level act on a round-capped, operator-owned
  document — the same boundary the S-H addendum respected when it routed spec findings through
  DECISION-NEEDED #4 rather than editing. Exact sites named so the gap is actionable.
• .claude/orchestrator-prompts/arch-v2-context-pipeline/calibration.md — the live ledger.
  :43 («Item 2 — ADR-8 window + owner») and :52 (the load-bearing baseline-capture boundary)
  still record the pre-descope owner. Deliberately not edited here: §6 names «the umbrella's
  ledger-schema section» as a surface S-K needs, and re-owning the ledger mid-PR is the
  scope-crossing Option B was rejected for.

NOT SWEPT, by ownership — and verified to need nothing (3):

• .claude/orchestrator-prompts/arch-v2-context-pipeline-s-b/kickoff.md (:109, :121, :194,
  :243) — closed stage; its kickoff is the historical record of what S-B was asked. Note in
  passing: :243 («No ADR-8 baseline rows — S-D captures those») was already stale before
  this PR, since S-D is CLOSED-NULL. Pre-existing, not created here, and not mine to rewrite.
• .claude/orchestrator-prompts/arch-v2-context-pipeline-s-c/kickoff.md (:11, :58, :175,
  :227) — reads ADR-8 for why metered seats matter and defines a Metered? column. ADR-8 still
  exists; only the A/B's host moved. Nothing to sweep.
• docs/meta-factory/research-patches/2026-07-31-l2-channel-verdict.md (:8, :16, :18-19) —
  append-only research patch; its ADR-8 references are Metered? (ADR-8) column semantics,
  untouched by a host change.

NOT SWEPT, correction already recorded elsewhere (1):

• docs/meta-factory/research-patches/2026-08-07-s-h-turn-attribution-p3d-p11.md:100-101 — fork
  feat: audit fixes Phase 1+2+3 (R2/R11 drift, manifest SSOT, depcruise integration) #1's Option B names «ADR-8's 20-dispatch window» as one of two baseline consumers. Under §5-A
  that consumer moves to S-K. The patch is append-only and read-only for later sessions, so the
  correction lives in the consuming artefact instead: docs(arch-v2): record operator verdicts on S-H DECISION-NEEDED #1 and #2 #1250 put it in the S-D′ kickoff's INPUT
  CONDITION footnote («one consumer, not two»), and §6 here gives S-K the same inherited condition.

Verified to have nothing to sweep (2): .claude/skills/dispatcher/SKILL.md and install.sh —
grep -c 'ADR-8\|parity(task-id)' → 0 on both. The A/B was never implemented on either surface;
that unimplementability is finding (iv) and half the reason for Option A. Both are named in §6 as
surfaces S-K will need.

Two surfaces are left with named, routed gaps (the spec and the ledger); every other surface is
swept, verified empty, or corrected in its consumer.