Decide whether the park path gets a composer, in a record superseding or extending ADR 0066

[randomparity]

Part of #380

Problem

#380 asks for a park-path composer for the WORK:TRAJECTORY block. An accepted decision record
already considered that exposure and declined to close it, so the first deliverable here is a
decision, not code.

ADR 0066 (docs/adr/0066-compose-the-handoff-handshake.md, Accepted 2026-09-06) records the park
exposure in its Consequences as "a real residual, not an oversight", and rejects two of the shapes
#380 gestures at. Anatomy rule 2 compounds this: the evidentiary ground for shipping
publish-handoff was four logged failures (issue #308). The park path has zero observed
failures, by #380's own admission — so rule 2's "performed inconsistently" limb is not available
here on the terms 0066 used it.

That does not make #380 invalid. An accepted ADR settles whether an exposure was known and
accepted, not whether it must stay accepted, and the operator's filing is new input the record did
not have. It does mean the outcome of this issue is genuinely open in both directions.

Explicitly in scope: concluding that the park path should stay prose, and closing the two
child issues that depend on this one. That outcome is a legitimate result of this issue, not a
failure of it.

Evidence

The accepted decision, and the three positions it already took:

• docs/adr/0066-compose-the-handoff-handshake.md:5 — Status Accepted (2026-09-06).
• docs/adr/0066-compose-the-handoff-handshake.md:34 — "The park path is unchanged and stays
  prose."
• docs/adr/0066-compose-the-handoff-handshake.md:125-129 — Consequences: "Park notes keep the
  sentinel-omission exposure the hand-off just lost. That is a real residual, not an oversight",
  with its stated recovery: a reader "recovers from by opening the issue".
• docs/adr/0066-compose-the-handoff-handshake.md:171-173 — Rejected, a
  --kind hand-off|park mode flag: "A park note that acquired a handshake would authorize a merge
  of work that is parked — strictly worse than the defect being fixed."
• docs/adr/0066-compose-the-handoff-handshake.md:165-170 — Rejected, a general
  post-annotation helper covering every WORK:* type.
• docs/adr/0066-compose-the-handoff-handshake.md:86-91 — the rule-2 argument the shipped helper
  stood on: "the evidence is the four holds issue Hand-off annotation is composed by hand and silently fails the merge gate #308 logs, from three different agents".

The three behaviours a park composer would have to differ on:

• skills/return-to-town/scripts/publish-handoff:22-23 — the two markers are hardcoded constants;
  compose_body at :403 always emits the handshake line, and validate_notes at :223 refuses a
  narrative carrying either marker or MERGE-READY:. There is no mode.
• skills/return-to-town/scripts/publish-handoff:342-382 — resolve_head requires GitHub's
  headRefOid to agree with git ls-remote origin, and refuses when they differ.
• skills/quest/SKILL.md:1052-1065 — the park operation is two steps in one order: post the
  WORK:TRAJECTORY note, then swap the status: label. publish-handoff touches no labels.

Expected

A record in docs/adr/ that supersedes or extends ADR 0066 and settles, on evidence, four
questions — with "leave the park path as prose" available as the answer to all of them:

1. Shape. A shared helper inside skills/return-to-town/scripts/publish-handoff versus a
   separate sibling executable. 0066 rejected the --kind flag; whether that rejection also
   forecloses a shared helper with separate entry points is this record's to decide.
2. The absent-handshake mode. A park note must not be able to acquire a MERGE-READY line.
   That is 0066's stated reason for rejecting the flag, and any shape chosen here has to make it
   unreachable by construction, not by caller discipline.
3. The relaxed branch-state refusal. publish-handoff refuses on remote-tip/headRefOid
   disagreement. That is correct for a hand-off and wrong for a park — a moved branch is a thing
   you might be parking about.
4. Label ownership. Who owns the status: swap, given that the label change is part of the
   park operation and the hand-off composer touches no labels.

The record must re-argue anatomy rule 2 without the four-failure evidence, since the park path
has none. If that argument cannot be made honestly, the record says so and the park path stays
prose.

Proposed approach

Write the record through $spellcraft, weighing the alternatives 0066 already rejected and stating
which of its judgments survive the new input and which do not. Cite 0066 rather than restate it.
just records runs the record gate; the version bump rules in CLAUDE.md apply.

If the verdict is "stay prose", close this issue by merging the record and close the two dependent
children as not planned, citing it.

Out of scope

These belong elsewhere and must not be reopened here:

• quest-log's missing MERGE-READY discriminator → issue merge-gate part 4 delegates selection to a quest-log recipe that lacks its MERGE-READY discriminator #378 (open, in flight).
• Timeouts on publish-handoff's git/gh calls → issue publish-handoff's git and gh calls carry no timeout, so an unreachable remote hangs an unattended worker #379 (open).
• Stating the closing-sentinel requirement in $return-to-town prose → issue return-to-town never states the closing-sentinel requirement its hand-off is judged against #375 (closed).
• Hand-off refusal behaviour — fork pull requests, closingIssuesReferences binding, remote-tip
  agreement for the hand-off → ADR 0066, unchanged.
• Composers for other WORK:* annotation types → rejected by ADR 0066 and not revisited.
• PR-side WORK:TRAJECTORY blocks in $return-to-town and $restock — proposed excluded, to keep
  file scopes disjoint across this decomposition.

Sequencing

Nothing blocks this issue. It blocks the other two children of #380: the composer's shape is this
record's to decide, and if the record concludes "stay prose" neither of them is built.

[randomparity]

annotation token: q381-14626d1f

scope identity: #381 — token q381-14626d1f

outcome: One new record under docs/adr/ that extends or supersedes ADR 0066 and settles, on
evidence, whether the park path gets a composer — deciding the issue's four questions (shape, the
absent-handshake mode, the relaxed branch-state refusal, label ownership) and re-arguing CLAUDE.md
anatomy rule 2 without ADR 0066's four-failure evidence. "The park path stays prose" is an
explicitly legitimate answer to all four.

completion criteria:

1. A record in docs/adr/ extending or superseding ADR 0066. Source: issue Decide whether the park path gets a composer, in a record superseding or extending ADR 0066 #381 Expected.
2. The record settles question 1, shape — shared helper inside publish-handoff vs. separate
   sibling executable, and whether 0066's --kind rejection forecloses a shared helper with
   separate entry points. Source: issue Decide whether the park path gets a composer, in a record superseding or extending ADR 0066 #381 Expected 1.
3. The record settles question 2, the absent-handshake mode, unreachable by construction rather
   than by caller discipline. Source: issue Decide whether the park path gets a composer, in a record superseding or extending ADR 0066 #381 Expected 2.
4. The record settles question 3, the relaxed branch-state refusal. Source: issue Decide whether the park path gets a composer, in a record superseding or extending ADR 0066 #381 Expected 3.
5. The record settles question 4, status: label ownership. Source: issue Decide whether the park path gets a composer, in a record superseding or extending ADR 0066 #381 Expected 4.
6. The record re-argues anatomy rule 2 without the four-failure evidence, and says so plainly if
   that argument cannot be made honestly. Source: issue Decide whether the park path gets a composer, in a record superseding or extending ADR 0066 #381 Expected, closing paragraph.
7. The record cites ADR 0066 rather than restating it, and states which of its judgments survive
   the new input and which do not. Source: issue Decide whether the park path gets a composer, in a record superseding or extending ADR 0066 #381 Proposed approach.
8. just records passes (adr profile). Source: issue Decide whether the park path gets a composer, in a record superseding or extending ADR 0066 #381 Proposed approach; CLAUDE.md
   Verifying a change.
9. .claude-plugin/plugin.json version is bumped. Source: CLAUDE.md / ADR 0022.

provenance:

• Outcome and all four questions: issue Decide whether the park path gets a composer, in a record superseding or extending ADR 0066 #381 body (Problem, Evidence, Expected).
• Parent direction and the asymmetry argument: issue The park path still composes its WORK:TRAJECTORY block by hand, keeping the exposure the hand-off path lost #380 body (native parent of Decide whether the park path gets a composer, in a record superseding or extending ADR 0066 #381).
• Prior accepted decision under re-argument: docs/adr/0066-compose-the-handoff-handshake.md.
• Observed-failure evidence weighed against anatomy rule 2's second limb: issue Hand-written tracking blocks reliably miss the closing COMPLETE sentinel #390 body — a
  post-merge audit finding all five hand-written blocks on one issue missing the closing sentinel,
  the one programmatically-emitted block correct, with a second worker's independent audit
  reproducing it. Supplied by the dispatching orchestrator as a verified correction to Decide whether the park path gets a composer, in a record superseding or extending ADR 0066 #381's
  "zero observed failures" premise, and revalidated against Hand-written tracking blocks reliably miss the closing COMPLETE sentinel #390 live at this step. It is evidence
  to weigh, not a verdict.
• Assigned ADR number (0069) and the reserved version value: campaign orchestrator dispatch.
• Exclusion approval: operator-confirmed 2026-09-16 at the campaign's step 4 plan table, relayed by
  the dispatching orchestrator; rechecked at this step against issue Decide whether the park path gets a composer, in a record superseding or extending ADR 0066 #381's own Out of scope
  list, which carries the same six exclusion/owner pairs unchanged.
• Interaction value: the dispatching orchestrator states no human is reachable for this run.

exclusions:

• quest-log's missing MERGE-READY discriminator — owner: issue merge-gate part 4 delegates selection to a quest-log recipe that lacks its MERGE-READY discriminator #378.
• Timeouts on publish-handoff's git/gh calls — owner: issue publish-handoff's git and gh calls carry no timeout, so an unreachable remote hangs an unattended worker #379.
• Closing-sentinel prose in $return-to-town — owner: issue return-to-town never states the closing-sentinel requirement its hand-off is judged against #375.
• Hand-off refusal behaviour (fork pull requests, closingIssuesReferences binding, remote-tip
  agreement for the hand-off) — owner: ADR 0066, unchanged.
• Composers for other WORK:* annotation types — owner: ADR 0066, rejected there and not revisited.
• PR-side WORK:TRAJECTORY blocks in $return-to-town and $restock — owner: none; excluded to
  keep file scopes disjoint across this decomposition.
• The hand-write instruction-shape defect itself — owner: issue Hand-written tracking blocks reliably miss the closing COMPLETE sentinel #390, queued in this campaign and
  merging after this row. This record decides the composer question only.

surface: one new file under docs/adr/ (number 0069, assigned); one light-spec artifact under
docs/workflow/specs/; the mandatory .claude-plugin/plugin.json version edit. No index row in
docs/adr/README.md — that directory deliberately carries no table and the adr record profile warns
W-INDEX-TABLE when one appears. Direct dependencies read but not changed:
docs/adr/0066-compose-the-handoff-handshake.md,
skills/return-to-town/scripts/publish-handoff, skills/quest/SKILL.md On a Blocker,
skills/quest-log/SKILL.md, skills/resurrection/SKILL.md, references/merge-gate.md.

ambiguities: none.

interaction: unattended

---

tracking metadata (not charter authority)

• blast radius: one new file under docs/adr/; no contract change; no code path altered
• change hazards: none
• complexity: S
• decompose verdict: one-pr
• routed review depth: single-pass
• classification: non-trivial
• artifact lane: light-spec
• design denominator: 100 changed lines (S) — derived live from issue Decide whether the park path gets a composer, in a record superseding or extending ADR 0066 #381's body (one record file,
  no contract change, docs/adr convention) and the cited ADR 0066 sections; issue Decide whether the park path gets a composer, in a record superseding or extending ADR 0066 #381 carries no
  WORK:DIVINATION block and no comments, so no persisted assessment was available to validate and
  all four fields were derived live. Recomputed independently and equal to the orchestrator's
  comparison value.

[randomparity]

scope token: q381-14626d1f

parked phase: $quest step 6 — adversarial branch review. The review loop stopped as blocked at
its iteration budget with a residual blocking figure of 1.

branch: feat/park-composer-decision-381 (4 commits, not pushed; no pull request exists)

guardrail status: green. just verify exit 0 (28 suites), just records exit 0,
scripts/check-public-safety.sh exit 0, scripts/check-plugin-version.sh exit 0 against the base
ref. Both focused contracts in the spec's Validation inventory were observed red and green.

what a human must decide

The record's verdict — the park path stays prose — has held through five review rounds, but its
argument has been refuted on a different load-bearing ground at almost every pass, and the
confirming pass raised a new one the budget leaves unfixed:

The rule-2 re-argument says the defect "tracks hand composition rather than any one type". Under
CLAUDE.md anatomy rule 2 as written, that clears the bar for any hand-composed block — a park one
included. The record then denies the bar using a retrospective-coverage test ("a park composer
would have corrected none of the audited five") that rule 2 does not state.

The finding is correct. It is not rejectable and not deferrable: it attacks completion criterion 6,
which is the record's central argument.

The remedy is known and small — concede that rule 2 is met, and rest the decision on CLAUDE.md's
Cut before adding ladder plus issue #390's ownership of the shared hand-write path, which reaches
every annotation type for less than a park-only composer reaches one. That reframe changes the
record's grounds, not its verdict.

It was not applied, because applying a fix on the final budgeted pass ships it unreviewed. Two
routes are open, and both are the operator's:

1. Approve continuation past the budget stop. $quest step 6 defines the resume: a fresh
   WORK:TRAJECTORY recording the approval, swap back to status:in-review, apply the reframe,
   and resume at step 7.
2. Re-scope or close. If the repeated reconstruction of the argument is itself the signal, the
   operator may prefer to reconsider whether this record should ship in this form.

three notes outstanding at the exit, none blocking: the ordinary-park residual is understated
(quest-log's latest-complete-wins returns a stale earlier block as current, rather than merely
losing the phase note, so the failure model's accepted reason is half-wrong); question 1's "two
entry points remove the flag" does not follow from ADR 0066's actual reachability ground; and the
sibling-executable verified: bullet's park-applicable line figure is unsourced.

one follow-up candidate, adjacent and owned elsewhere: the reserved manifest version conflicts
with a sibling campaign row if this row merges first. That is the campaign orchestrator's merge
ordering, and the reserved value is not this run's to change.

review rounds: 5 this run, 5/1 cumulative for this change — two design-artifact passes, one
routed single-pass branch review that escalated on three blocking findings, and two $trial-loop
iterations. No handshake is emitted: this is a park, not a hand-off.

[randomparity]

scope token: q381-14626d1f — the frozen WORK:SCOPE charter identity, unchanged. This resume
does not re-freeze the charter. The claim lease released at the park was replaced with a fresh
token on resume, as authorized; claim recovery was not used.

phase: resuming $quest step 6 via the approved-continuation path, from the $trial-loop
budget stop recorded in the previous WORK:TRAJECTORY.

authorization: the operator, explicitly and on the record, relayed by the campaign orchestrator
in this run's resume dispatch. The orchestrator states it verified both findings against the record
itself before taking them to the operator rather than relaying them unchecked.

what it authorizes — exactly five items, and no widened scope:

1. Reframe the rule-2 argument: concede the performed inconsistently limb is met, and rest the
   decision on CLAUDE.md's Cut before adding ladder plus issue Hand-written tracking blocks reliably miss the closing COMPLETE sentinel #390's ownership of the shared
   hand-write path. The retrospective-coverage test is deleted rather than reworded — a proposed
   script is not measured by how many past instances it would have caught. The verdict does not
   move: the park path stays prose.
2. Replace the unsourced park-applicable line figure with a reproducible one.
3. Correct the ordinary-park residual: quest-log's latest-complete-wins returns a stale earlier
   block as current, rather than merely losing the phase note.
4. Fix or drop question 1's "two entry points remove the flag" inference.
5. One confirming pass with a fresh reviewer, independent of the one that raised the blocking
   finding.

explicitly not authorized: any further scope, and any seventh round. A new blocking finding
outside these five parks the run again rather than being fixed.

base refresh: main advanced to 1d8d4fe while this issue was parked (wave 1 drained). The
branch takes that base by merge, not rebase, and the .claude-plugin/plugin.json conflict
resolves by keeping this row's reserved 5.11.2, which remains strictly greater than the
5.11.1 now on main. No reassignment was issued.

resolved without filing: the previous park's single follow-up candidate — the reserved version
versus the sibling row — was a merge-ordering question, and the ordering held.

review rounds: 5 spent before this authorization, 5/1 cumulative. This continuation adds the
confirming pass as round 6.

No handshake is emitted here: this block records a resume, not a hand-off.