D: --deep and Phase 6 boundaries (F-457), guard repeat-ask remedy (F-453), work-ordering bullet (#1) — plugin 0.41.1

[BradleyDB]

Base branch must be dev, not main. Base is dev.

Regenerated output committed? build/build-plugin-gs-superadmin.mjs was run; nothing under data/, reference/, wiki/ or plugins/gs-superadmin/reference/ changed (prose + one hook message only).

What

Session D of the September rounds: DC-1 (bus F-457), DC-2 (F-453), DC-3 (issue #1). Plugin 0.41.1.

• --deep <domain> states its phases (setup/SKILL.md): Phases 1–2, then straight to Phase 5 — Phases 3 and 4 never re-run and nothing re-lists. Ruled 2026-09-16 over the plan's re-list-when-stale option after measuring the selection: next --upgrade takes depth-metadata entries only, so a re-list's new pending rows would never enter the --deep queue. Before spending budget the run reads one report's lookback.<domain> and relays the list's age when older than lookbackDefault, pointing at refresh (refresh step 2 stays the one home for the window). Phase 4's heading no longer reads "runs every time".
• Phase 6 states the maps' inputs: the five lanes are the builder's ONLY inputs — the subset it takes from doc-lib's RECORDED_LANES, pointed at, not copied — so a non-lane deep ingest leaves the four maps byte-identical; deps-report is named as where that ingest pays off.
• The guard's repeat-offense ask names the remedy (hooks/gs-admin-guard.mjs): the escalated ask on a shell-variable subcommand carries the same "spelled literally — variables belong in flag values and paths" remedy the deny already carried. test/guard-fixtures.mjs pins it on the repeat path.
• AGENTS.md § Review-gate rules gains "Work order is not severity order" (marked not mechanically checkable). The same rule lands in the /dev-loop skill's builder lane by its own PR, BradleyDB/gs-admin-superfriends#8 — the personal skill is a symlink into that checkout.

Bus: F-457 and F-453 FIXED with Fix / Judge lines; the D-V walk (one setup --deep run to its first batch + the repeat ask read by the operator) is banked in dev/VALIDATION.md.

Closes #1

Why

F-457: four live --deep runs each improvised the same skip of a documented "runs every time", and several login cycles went to deep-ingesting a non-lane domain to improve maps that came back byte-identical — two expensive operations documented by what they do, never by what they do not. F-453 (polish, carried since Session A): the rung-1 site present at the moment of composition blocked correctly and did not say how to not need the prompt. #1: the ordering rule learned at 0.30.0 was lost with the private bus.

Review notes

Start with the Phase 5 --deep paragraph (the canon; the flag doc and Phase 4 heading are paraphrase + pointer), then the Phase 6 lane paragraph, then the varAskPart string. Mutation proof: guard-fixtures over the stashed pre-fix hook reds exactly the new check; the F-363 bound arm measures 1967 of 4096 bytes. Locally green: check-doc-drift, check-stale-facts, check-imports, check-instance-data, test-check-doc-drift, guard-fixtures, guard-oracle, claude plugin validate, plugin bundle rebuild (no diff), sweep-fence-grammar (survivor comparison EXACT). Review round (/code-review low): 1 finding (the null-days substitution rule produced an ungrammatical relay line), fixed.

• Plugin version bumped in .claude-plugin/plugin.json with a CHANGELOG entry (0.41.1 — the ask text ships).
• Safety boundary: this PR touches hooks/gs-admin-guard.mjs — one message string only; no ask became a deny or disappeared (the deny text is unchanged, the ask gains a sentence); test/guard-fixtures.mjs extended and mutation-proved.
• Skill edits: cross-document consistency checked (the flag doc, Phase 4 heading and Phase 5 canon agree; the operating model and cheatsheet carry no --deep phase claim) — the setup walk is the tester round's. Session / verdict: D-V, per dev/VALIDATION.md § F-457 / F-453.
• No test or dev step touched real gs-admin state.
• No tenant/instance/org-specific data in commit messages or this PR body.