Step 10d: payload-aware no-PR arm (a branch with novel payload and zero PRs silently skips the auto-merge)

Goal

Close a silent payload-stranding gap in /issue Step 10d: when a branch has no PR object at all, the auto-merge takes an unconditional skip arm that never consults the novel-payload predicate sitting a few lines below it — so a code-bearing branch is left permanently unmerged with the durable record reading clean.

The gap (hit live on #2235, 2026-08-11)

Step 10d's safe-case procedure opens with the PR-object liveness probe (.claude/skills/issue/SKILL.md:12698-12704):

PR=$(echo "$PR_INFO" | cut -d' ' -f1)
...
if [ -z "$PR" ]; then
  echo "No PR for issue-<N>; nothing to merge."   # skip; post nothing

The -z "$PR" arm is unconditional: it prints "nothing to merge" and skips, posting nothing. It never asks whether the branch actually carries payload.

Contrast the very next arm (PR_STATE != OPEN, lines 12706-12751), added by #1897 for exactly this hazard: when a terminal PR exists, the step runs the layered novel-payload predicate (commits-ahead → git cherry patch-ids → content-identity vs origin/main) and, on NOVEL_PAYLOAD=yes, creates a fresh draft PR and merges it. So the workflow already knows that "no usable PR object" plus "novel payload" must resolve to create-then-merge, not skip — it just does not apply that knowledge when the PR count is zero rather than one-and-terminal.

Observed on #2235. gh pr list --head issue-2235 --state all returned [] — no PR had ever been created (the draft PR that normally lands at the approved→running transition never happened; that upstream miss is the second half of this bug). The branch was 15 commits ahead with 14 git cherry + lines, carrying scripts/workflow_lint.py, .claude/skills/issue/SKILL.md, scripts/inline_lint_gate.py and three test files. Left to the skip arm, all six deliverables — including two SHARED workflow-surface files — would have stranded on an unmerged branch while the task transitioned to completed.

Why that is the expensive class. This is the #456 → #466 shape Step 10d was built to prevent, and its own rationale says so (SKILL.md:11005-11010): a shared trainer.py fix stranded on a deferred branch, and the next experiment inheriting from main lacked it and crashed. A stranded workflow_lint.py is worse than a stranded experiment helper, because the no-flags lint IS every session's Step 9c instrument — the #1388 fleet-wide coupling.

Note the existing detectors do NOT cover it: the watcher's completed_unmerged_pass (#1540/#1653) keys on a completed task with no epm:merged marker, but the skip arm posts no marker at all and #1723 routes the completed flip through Step 10d's terminal-teardown — so on the skip arm the task may never reach completed via that path, and the pass has nothing to latch onto. Silence on both sides.

Fix

Two independent changes; (1) is the load-bearing one.

(1) Make the -z "$PR" arm payload-aware — reuse the #1897 predicate verbatim. Hoist the layered novel-payload predicate above the PR-state branch and gate BOTH no-usable-PR arms on it:

• NOVEL_PAYLOAD=no → keep today's skip ("nothing to merge"), post nothing.
• NOVEL_PAYLOAD=yes → take the #1897 fresh-PR path (gh pr create --draft, re-resolve PR_INFO, then guards + merge). The draft must be marked ready (gh pr ready) or gh pr merge refuses — see (3).

The predicate is already written, already fail-SAFE toward "novel" on every git-error path, and already trusted for the terminal-PR case; this is a reuse, not a new mechanism.

(2) Fail loud when the upstream draft-PR creation never happened. The state machine (SKILL.md § The State Machine) shows approved |-- (worktree + draft PR) |--> running, but nothing verifies the PR exists afterward. Add a cheap assertion at the first step that needs it (Step 5's round-push, or Step 10d entry): a branch at/past running with commits ahead of origin/main and zero PRs is an anomaly worth one loud line, since it means an earlier transition silently no-opped.

(3) Document the draft-PR merge precondition. A --draft PR is not mergeable; gh pr ready must run first. The #1897 fresh-PR arm creates a draft and proceeds straight to the merge block with no gh pr ready between them, so that arm is latently broken for the same reason — worth fixing in the same pass.

Acceptance

• A branch with novel payload and ZERO PRs gets a PR created and merged by Step 10d (no silent skip).
• A branch with NO novel payload and zero PRs still skips, posting nothing (today's behaviour preserved).
• The #1897 fresh-PR arm marks the PR ready before merging.
• A test pins the payload-aware routing of the -z "$PR" arm.

target_file: .claude/skills/issue/SKILL.md problem: Step 10d's if [ -z "$PR" ] arm (SKILL.md:12703) skips the auto-merge unconditionally when no PR object exists, without consulting the layered novel-payload predicate used a few lines later for terminal PRs (#1897). A code-bearing branch whose draft PR was never created is therefore left permanently unmerged with no marker posted — the #456/#466 stranded-shared-module class. Hit live on #2235: zero PRs, 15 commits ahead, 14 cherry '+' lines, payload including scripts/workflow_lint.py and .claude/skills/issue/SKILL.md. The watcher's completed_unmerged_pass cannot catch it because the skip arm posts no marker. fix: hoist the novel-payload predicate above the PR-state branch and gate the -z "$PR" arm on it — NOVEL_PAYLOAD=no keeps today's skip, NOVEL_PAYLOAD=yes takes the existing #1897 fresh-PR create-then-merge path; add gh pr ready before the merge on both fresh-PR arms (a draft PR is not mergeable); add a loud assertion when a branch at/past running has commits ahead and zero PRs. urgency: normal wf_fix: true confidence: high related_task: #2235

Provenance

Hit during #2235's Step 10d auto-merge on 2026-08-11. Not a speculative read of the recipe: the PR probe returned empty, gh pr list --head issue-2235 --state all returned [], and the novel-payload predicate returned 15 commits ahead / 14 cherry + lines. The merge was completed by creating the missing PR manually (#1912); this task makes that recovery automatic.