docs(deps-hash): night-mode execution prompt for umbrella S2→S3 sequenced closure

[artyhoo]

What

Single self-contained prompt for a fresh Claude Code night-mode session to finish the deps-hash-multistack umbrella (kickoff #1016). DH-S1 is LANDED on staging (#1024 + #1029 merged); this prompt sequences the remaining DH-S2 (rust detection) → DH-S3 (closure) with all the discipline + learned incidents encoded.

Why a written prompt (not inline chat)

Umbrella = 2 sequenced stages + 1 open maintainer-fork (S2 size-gate A-vs-B) + 3 decision points + recursive STOP-lines. Holding this in fresh-session memory = fact loss → exactly the drift the project exists to prevent. The file is the single source of truth.

What the prompt guarantees

• Correct order: S1 done → /pipeline deps-hash-multistack → S2 → stage-gate → S3 → done.md.
• Fork parked, not silently decided: S2 size-gate (hook already 212 lines on staging; +rust Tier-1 ~+30-40 → ~240-250, over the ~180 soft threshold) → AskUserQuestion to maintainer, not autonomous pick.
• All kickoff §2 STOP-lines enumerated: no-lib delivery, _emit_warn preserved, always-exit-0, @dual-pair same-commit, no TOML-parser dep, no setup.d/NN-rust.sh, done.md only at DH-S3.
• Iterative dual cold-context review per increment — the method that caught round-3.5 BLOCKERs (CRLF, python-upgrade drift, fresh-install-python) invisible to the author + 23-test suite.
• Learned incidents encoded as warnings: local-ref staleness (gh pr diff is truth, not local git diff), flaky-CI rerun (single job via gh api), parallel-checkout stash, untracked §1.7-breaking research-patches parking, LF-only fixture blind spot.

What it does NOT guarantee (by design)

• The S2 fork outcome (A=one-hook vs B=split) — decided by measured verdict + maintainer answer.
• New edge-cases won't appear — night-mode autonomously parks + continues unambiguous parts.

Verification

• markdownlint: 0 errors.
• The doc lives at docs/meta-factory/ (NOT research-patches/), so principles/13 §1.7 does not apply.
• All facts re-verified against origin/staging at write time (2026-07-18): hook = 212 lines, DH-S1 fully landed (feat(deps-hash): DH-S1 multistack — JS-widen + python Tier-1/2 + per-stack baselines (#1016) #1024 + fix(deps-hash): DH-S1 round-3.5 cold-context review — CRLF + python3-upgrade + fresh-install (follow-up to #1024) #1029), done.md absent.

§1.7 Backward-check

Class of this change = author a working prompt doc for sequenced stage execution. Surfaces where this class must also hold:

1. The kickoff itself (already documents S2/S3 deliverables) — SWEPT-CLEAN, the prompt references §2 stages verbatim, does not contradict.
2. The night-mode skill (referenced as the executor) — SWEPT-CLEAN, the prompt's autonomy policy matches night-mode delta item 1 (technical fork → resolve with rationale; owner fork → park).
3. The fresh CC session that will consume it — GAP-FOUND-BY-DESIGN: the session has zero context about this umbrella's history; the prompt is the bridge. This is the intended design, not a gap to close.

§1.7 Forward-check

Complies with doc-authority-hierarchy.md (the prompt carries a scope note + is filed under docs/meta-factory/ as a working artifact, not a canonical rule); build-first-reuse-default.md (references /pipeline + night-mode skills rather than re-describing their machinery); no-paid-llm-in-ci.md (the prompt runs in a session-bound CC subscription, zero API calls); ai-laziness-traps.md T3 (every state claim cites a PR number or file:line the fresh session can verify).

Prior-art: extends the project's existing pipeline/night-mode skill substrate (no new capability — a working prompt that orchestrates them for one umbrella's closure).