feat(beta-delivery-ux): S2 — pipeline presets + status + workspace one-command, with the refresh-drift gate closed

[artyhoo]

Ships S2 (pipeline launch presets + /pipeline status + workspace one-command, spec A4/A5/A9) and closes the three items that left it HALTED at the §2.4 two-REVISE cap.

1 — do_refresh() omitted the four worktree scripts (the RED gate on this branch)

setup.d/85-worktree-scripts.sh:50 copy_safes create-worktree.sh, worktree-node-modules.sh, link-coordination.sh and getff-work.sh on the --full env+ path, with no refresh arm. A brownfield env+ consumer therefore received none of Park-6's «full functional set» on --refresh — the #869 class this gate exists to catch. Verified against the merge-base rather than assumed: 7d44221af3 = 6/6 PASS, branch = 5/6.

The list has to be duplicated — the delivery site is a setup.d module sourced only during install, so do_refresh cannot read its array — and the set-difference check cannot see a per-script divergence: both sides collapse to the token scripts/, so adding a 5th script to the delivery array while forgetting the refresh loop would stay GREEN. Added a parity check that compares the two lists directly, plus its non-vacuity negative.

Falsified for real, not asserted — a 5th name added to the delivery array only:

✗ worktree-scripts parity: delivery/refresh lists DIVERGE —
  delivered=[brand-new-thing.sh create-worktree.sh getff-work.sh link-coordination.sh worktree-node-modules.sh]
  refreshed=[create-worktree.sh getff-work.sh link-coordination.sh worktree-node-modules.sh]
PASS=7 FAIL=1        → restored: PASS=8 FAIL=0

2 — render-status.sh:104 did not compile, and the fallback hid it

The → and the mergeable= label sat inside the \(…) interpolations, which is a jq syntax error; the 2>/dev/null || echo fallback on the next line swallowed it, so §3's «open PRs + mergeable state» rendered the fallback line on every single run. Live proof:

old → jq: error: syntax error, unexpected INVALID_CHARACTER (exit 3)
new →   - #42 Test [feat → main, mergeable=MERGEABLE]        (exit 0)

3 — it.fails() is no longer a delivery mechanism

NON-DEGRADED-PR-SECTION shipped as it.fails() with the defect written up as an operator TODO in the impl-note §8.2 — a known-broken path presented as a passing suite. That is exactly the #warning-nobody-reads shape attention-is-not-a-mechanism.md §1 rejects. With the template fixed the marker is gone and the guard is live: re-introducing the defect → 1 failed | 5 passed; restored → 6 passed.

4 — SKILL.md §3 Step 3b, the last open MAJOR

Spec A4 (lines 269-271) TTY preset row, additive and TTY-gated; --preset / AIF_PIPELINE_PRESET stays primary per kickoff §2 binding constraint 1, and menu-only UX stays REJECTED. The block lives in .claude/skills/pipeline/references/output-format.md:493 (§8) rather than inline, because inlining pushed SKILL.md to 602 lines and tripped the 600-line markdown gate; SKILL.md:309 keeps a one-line pointer. Principle 18 still 6/6.

On the «permission-blocked» claim

Items 2 and 4 were reported across both rework rounds as «PERMISSION-BLOCKED — .claude/skills/pipeline/ not writable». The cold auditor read that as contradicted by a diff that creates files under exactly that path. Both readings are wrong, in different directions.

The auditor's contradiction does not hold: the two rework commits (dfc955842b, 6564566d67) touch .claude/skills/pipeline/ zero times, while the three initial commits wrote nine files under it. The claim was about the rework rounds, and the commit record is consistent with it.

But the stated mechanism does not exist on any surface reachable from here: the repo deny list carries no .claude/skills entry (.claude/settings.json), the container's global settings have permissions: null, all seven PostToolUse Edit hooks exit 0 on both paths, and the container filesystem is writable as root. So the claim is not falsified and not corroborated — and the practical consequence is the one that matters: both edits were made here without incident.

Verification

bash tests/install-sh/refresh-covers-full-delivery.test.sh   → PASS=8 FAIL=0
npx vitest run packages/core/hooks/render-status.test.ts
        packages/core/principles/18-meta-orchestrator-output-format.test.ts → 12 passed
bash tests/install-sh/glm-onebutton.test.sh (post-merge, S4 on staging)     → PASS=33 FAIL=0

All on the host. The refresh gate was re-run after merging staging (which brought S4's install.sh changes) and still reports 8/8, so no sibling delivery site regressed.

§1.7 Forward-check applied

Complies with attention-is-not-a-mechanism.md:29 — the it.fails() removal at packages/core/hooks/render-status.test.ts:105 converts a check whose failure mode was «nobody looked» into one that fails, and the new parity arm at tests/install-sh/refresh-covers-full-delivery.test.sh:1 is a deterministic gate at the earliest reachable channel (pre-push) carrying its own non-vacuity negative. Complies with no-paid-llm-in-ci.md:1 (awk/grep/jq only, zero API-billed calls). Complies with doc-authority-hierarchy.md:1 — the new §8 lands in a file that already carries a Class + Authoritative-for header, and the §3 pointer keeps SKILL.md under the 600-line ceiling rather than growing the hot doc. Complies with language-discipline.md:1 (machinery in English).

§1.7 Backward-check applied

Class of this change = a --full copy_safe delivery site with no matching do_refresh arm. Enumerated beyond this diff: setup.d/45-python.sh:1 — SWEPT-CLEAN, it owns its own GETFF_TOOLCHAIN_REFRESH path and Check 4 of this same gate already asserts that parity; setup.d/40-configs.sh:1 — SWEPT-CLEAN, its deliveries are either refreshed or in the EXCLUDED list, which Check 2 proves is not stale; setup.d/aif-handoff-guided-install.sh:1 — SWEPT-CLEAN, it is invoked from install.sh:1139 and is not a copy_safe delivery at all, so the class does not reach it. Second class = a shipped helper whose failure is masked by a 2>/dev/null || fallback: swept .claude/skills/pipeline/helpers/list-presets.sh:1 and .claude/skills/pipeline/helpers/resolve-preset.sh:1 — SWEPT-CLEAN, both carry paired-negative coverage (packages/core/hooks/list-presets.test.ts:1, packages/core/hooks/parse-preset.test.ts:1) that exercises the real jq path rather than only the fallback.

Prior-art: prior-art-evaluations.md#87 (Worktree-creation CLI landscape, verdict BUILD — scripts/getff-work.sh and setup.d/85-worktree-scripts.sh do NOT rebuild worktree creation; they compose the already-BUILT scripts/create-worktree.sh from that row and merely ship it to env+ consumers, which is the delivery half #87's BUILD verdict presumes).
Prior-art: prior-art-evaluations.md#110 (cross-worktree gitignored-SSOT symlink sync, verdict ADAPT — link-coordination.sh is the artefact that row owns; this PR adds it to the shipped set and to do_refresh, changing no behaviour of the script itself).
Prior-art: clig.dev flag-first, verdict REFERENCE, no new SSOT row — upstream problem class = non-interactive CLI UX; ours = preset activation must work for agents and CI, not only an interactive TTY. Match is identical, so the flag/env path stays primary and the §3 Step 3b TTY row is strictly additive.

Fidelity verdict

FIDELITY: skipped — no cold fidelity auditor was run in this session, so a GO would be unearned; this is a dispatcher-side repair of the S2 halt, self-reviewed on the host, with every claim carrying a command output or an in-repo file:line and two falsification runs (parity gate, render-status guard) standing as the non-vacuity evidence