docs: merge authority stays with the owner (OD-0020); record the unfinished test pruning

.agent/lessons/INDEX.md

@@ -1,5 +1,5 @@
 # Lessons — always-on index
-last-curated: 2026-07-17
+last-curated: 2026-07-29
 
 These fire every session (auto-loaded via CLAUDE.md). Topic bodies live in `topics/` — read a shard only when its pointer matches. Cap: 10 always-on lessons; when full, demote the least-actionable into a shard rather than deleting it.
 
@@ -10,5 +10,7 @@ These fire every session (auto-loaded via CLAUDE.md). Topic bodies live in `topi
 
 - `COMPLETED_WORK.md`'s "shipped" is not owner-verified-complete — never write a public wiki changelog entry or delete a source plan on its word alone (2026-07-18) — during the release doc reduction, COMPLETED_WORK listed HuggingFace voice, AI casting, and the recording-cue expansion as "shipped/live/all shipped." Acting on that, I wrote wiki changelog entries for two and marked their plans deletable. Owner review corrected all three: HF is untested/needs sign-off, AI casting is placeholder/future in the app, recording-cue awaits the owner's portrait-image generation. The changelog entries were reverted; no plan had been deleted (the ledger gate held). **Apply:** treat "shipped" in a summary doc as a claim to verify, not a fact. Before writing the history-of-record (wiki changelog) or deleting a plan, confirm completion against on-disk reality AND owner sign-off — "code + endpoints landed" ≠ "feature complete." A feature marked future/placeholder in the app, or awaiting an owner action, is not a changelog entry and its plan stays. The consolidation ledger (`design-docs/plans/active/final_release/21_*`) is the gate that made this recoverable — keep deletions gated on it.
 
+- `studio-2.0` squash-merges, so a branch's commit SHAs and subjects NEVER appear in its history even when the work fully landed (2026-07-29) — reconciling 32 abandoned branches, a first pass matched commit subjects against `git log origin/studio-2.0` and reported all 32 as unlanded, which was false for nearly all of them; `git rev-list --cherry-pick` was equally useless, because a squash rewrites the patch-id of the whole branch into one commit. Reachability (`--contains`, `merge-base --is-ancestor`) is also unreliable for the same reason, and inflates the count: N branches holding one shared squashed trunk each read as N distinct bodies of unlanded work. **Apply:** to answer "did this branch land", read what the commit changed and then read the CURRENT content of those same files on `origin/studio-2.0` — for a deletion, check the deleted file/test name is also absent there; for an edit, compare the file's content. Only content answers it. Corollary for the reverse direction: a branch whose files no longer exist on `studio-2.0` at all is usually obsolete-by-restructure, not unlanded. Worked verdicts for 32 branches: `.agent/reports/2026-07-29-reconcile-*.md`.
+
 ## Topic pointers (read on match)
 <!-- Add a line here when a lesson needs more than one sentence: "<trigger keywords> → topics/<name>.md" -->

.claude/agents/roster.json

@@ -192,7 +192,8 @@
       "audio-quality verdicts",
       "information-architecture calls",
       "release-facing defaults",
-      "merging a PR, cutting a release, posting outside this repo",
+      "merging a PR (not delegated while studio-2.0 is the staging target; a per-task grant expires with that task, OD-0020)",
+      "cutting a release, posting outside this repo",
       "amending CLAUDE.md's binding sections or the mandate"
     ]
   }

.claude/decisions/0020-merge-authority-stays-with-the-owner-until-studio-2-0-lands.md

@@ -0,0 +1,39 @@
+# OD-0020: Merge authority stays with the owner while `studio-2.0` is the staging target
+Status: accepted        Date: 2026-07-29
+Scope: `CLAUDE.md` (ask-first list); `.claude/agents/roster.json` (`unowned_surfaces.owner_only`)
+
+Context: On 2026-07-18 the owner did grant merge authority in conversation ("I give you authority to
+merge when you deem it ready"), and it was used once, on PR #163. That grant was recorded only in an
+untracked memory file; `CLAUDE.md` and `roster.json` were never updated and continued to list merging as
+ask-first, so the estate held two contradictory answers for eleven days. The 2026-07-29 retrofit surfaced
+the contradiction and the owner resolved it by **narrowing the grant**: he authorized merging for one
+specific task (landing PRs #174, #176, #177) and stated he retains the authority generally, because
+`studio-2.0` is a temporary staging line and he is particular about nothing merging to the wrong branch
+or being destructive while that is true.
+
+Decision: "Merging a PR, cutting a release, or posting anywhere outside this repo (issues/PR comments on
+someone else's thread, external services)" remains on the ask-first list. Opening a PR, pushing a branch,
+and choosing draft-versus-ready stay do-then-report; only the merge itself is withheld. A per-task
+authorization covers that task and expires with it: it is never read forward into a standing grant.
+
+Consequences: Narrows the real 2026-07-18 grant rather than denying it happened. Two lessons sit inside
+that: a grant recorded only in untracked memory produced a contradiction nothing could detect, which is
+the case for writing rulings into tracked files (OD-0002's closeout discipline), and a conversational
+grant can be narrowed later without either party having been wrong at the time. Blocks two specific
+reflexes: citing a past one-off
+authorization as ongoing permission, and merging a correct-looking change to the wrong base while
+`studio-2.0` is standing in for `main`. Note the asymmetry this preserves: Ada can still get work all the
+way to a reviewed, mergeable PR without asking, so the cost of the restriction is one message at the end,
+not a slower pipeline. This is also the reason the change that records this ruling could not be
+self-merged.
+
+**This expires by its own terms.** The owner stated he will loosen the mandate once `studio-2.0` merges
+into `main`. That event, not a date, is the trigger to revisit — and revisiting means asking him, not
+assuming the loosening is automatic.
+
+Disconfirming evidence: Merges are consistently routed to the owner as required, and the ask turns out to
+be pure latency — every request is approved unread over a sustained period, with no merge ever redirected,
+delayed, or rejected on inspection. That would show the gate is catching nothing the PR review already
+caught. Note what does **not** count: an agent merging anyway is evidence about the agent, not the rule;
+and a long quiet stretch with no wrong-base incident may be the rule working rather than the rule being
+unnecessary.

.claude/decisions/INDEX.md

@@ -28,3 +28,4 @@ the other.
 | [OD-0017](0017-orchestration-layer-edited-only-from-rooted-session.md) | accepted | CLAUDE.md orchestrator section; .claude/agents/ profiles; .claude/decisions/; roster.json; roster.html | This repo's orchestration layer is edited only from a session rooted in this repo | yes | — |
 | [OD-0018](0018-worktrees-are-not-the-default-and-are-removed-on-merge.md) | accepted | `CLAUDE.md` (Orchestrator identity & mandate); `.agent/rules/memory-queue.md`; agent dispatch decisions | Worktrees are not the default workflow, and a worktree is removed when its PR merges | yes | — |
 | [OD-0019](0019-a-functional-affordance-survives-a-redesign-only-if-defended.md) | accepted | `.claude/agents/junia-designer.md` (designer convictions); any redesign, visual migration, or screen-parity work | A functional affordance survives a redesign only if it is defended explicitly | yes | — |
+| [OD-0020](0020-merge-authority-stays-with-the-owner-until-studio-2-0-lands.md) | accepted | `CLAUDE.md` (ask-first list); `.claude/agents/roster.json` (`unowned_surfaces.owner_only`) | Merge authority stays with the owner while `studio-2.0` is the staging target | yes | — |

CLAUDE.md

@@ -56,7 +56,7 @@ Do-then-report (no permission needed):
 
 Ask-first (unchanged):
 
-- Merging a PR, cutting a release, or posting anywhere outside this repo (issues/PR comments on someone else's thread, external services).
+- Merging a PR, cutting a release, or posting anywhere outside this repo (issues/PR comments on someone else's thread, external services). **Merge authority is the owner's and is not delegated (OD-0020)** — `studio-2.0` is a temporary staging line, and the risk being managed is a correct-looking change landing on the wrong base. A per-task authorization to merge covers that task only and expires with it; do NOT read a past one-off grant forward as standing permission. Take the work all the way to a reviewed, mergeable PR without asking, then hand over the merge.
 - Destructive or hard-to-reverse operations: data deletion, schema migrations, contract version bumps beyond the task at hand, reversing an ADR.
 - Amending CLAUDE.md's binding sections or this mandate itself.
 - Perceptual and product judgment: audio-quality verdicts, information-architecture changes, release-facing defaults — stage the evidence (A/B samples, screenshots, diffs) for the owner; never assert the verdict.

design-docs/plans/REMAINING_TASKS.md

@@ -116,6 +116,22 @@ observation is missing.
       spec→provenance citation map: [active/final_release/20_stale_docs_retirement.md](active/final_release/20_stale_docs_retirement.md)**
       (Section A already done; Section B is the repoint-then-delete remainder).
 
+## Test-suite hygiene
+
+- [ ] **Finish the 2026-07-10 low-value test pruning (~40 tests still present).** The audit landed
+      (`active/final_release/audits/`) and most of its execution branches merged, but three never did:
+      Voices/VoiceLab (`751e6bd4`), backend DB (`7753b3fd`), and misc frontend pages (`f3493c36`), all
+      on `origin/studio2/phase-12.5-style`. The work is pushed and safe, just unmerged.
+      **Do not cherry-pick** — all three conflict in the same three VoiceLab test files
+      (`TestSection.test.tsx`, `VariantsSection.test.tsx`, `VoiceIconControls.test.tsx`), which have
+      moved on since. Redo the deletions against today's files instead, using each commit as the list
+      of what to remove and re-checking each against `testing-standards.md` R1–R4 before deleting
+      (a test that was low-value in July may have become load-bearing since). Rough cost: an hour.
+      Worth doing: the vitest suite leaks memory, so a smaller suite is a real win.
+      Evidence and per-branch verdicts: `.agent/reports/2026-07-29-reconcile-codeql-and-redundant.md`.
+- [ ] Fix the hardcoded-style violation at `frontend/src/pages/Voices/.../ArchetypeQuickPick.tsx:242`
+      (found in passing during the 2026-07-29 reconciliation; unrelated to the branches assessed).
+
 ---
 
 *When every item above clears and Stage 6's tag lands, this file should be empty — retire it and