diff --git a/GOALS.md b/GOALS.md
index 579c405d..5148df8a 100644
--- a/GOALS.md
+++ b/GOALS.md
@@ -6,8 +6,9 @@
 > line. The `Platform-Neutrality Rule` at the top of `CLAUDE.md` is the *operational
 > rule*; **this file is the source of truth for *intent* and *sequencing*.**
 >
-> _Last updated: 2026-09-03 — the queue drained; the mandated post-#735 audit is
-> now run and its numbers are below. Milestone #7 is **0 open**._
+> _Last updated: 2026-09-05 (issue #827 correction pass — the 2026-09-03 revision
+> went stale within 81 minutes of its own edit and stayed that way for two days;
+> see "#47 was closed without its blocker" below). Milestone #7 is **0 open**._
 
 ## 🎯 North star
 
@@ -83,15 +84,21 @@ closed.
 The previous revision of this file said: *"re-run the sizing audit when #735 lands
 rather than assuming the chain finishes the job."* Done, on `develop @ eedebf8`.
 
-Production lines, `#[cfg(test)]` excluded, all four columns measured with the same
+Production lines, `#[cfg(test)]` excluded, all columns measured with the same
 script (`scripts/prod_lines.py`) so they are comparable to each other:
 
-| | 2026-05-01 | 2026-07-01 | pre-chain 2026-08-31 | **now 2026-09-03** |
-|---|---|---|---|---|
-| `src/gtk/` | 18,969 | 13,675 | 12,526 | **9,650** |
-| `src/tui_main/` | 14,649 | 10,358 | 11,125 | **10,345** |
-| **both backends** | 33,618 | 24,033 | 23,651 | **19,995** |
-| `src/render.rs` (shared) | 10,574 | 12,807 | 15,009 | **21,405** |
+| | 2026-05-01 | 2026-07-01 | pre-chain 2026-08-31 | pre-#785 2026-09-03 | **post-#785, now @ `ee26268`** |
+|---|---|---|---|---|---|
+| `src/gtk/` | 18,969 | 13,675 | 12,526 | 9,650 | **2,607** |
+| `src/tui_main/` | 14,649 | 10,358 | 11,125 | 10,345 | **10,366** |
+| `src/app.rs` (hoisted out of `src/gtk/` by #785) | — | — | — | — | **7,131** |
+| **all three files** | 33,618 | 24,033 | 23,651 | 19,995 (2 files) | **20,104** |
+| `src/render.rs` (shared) | 10,574 | 12,807 | 15,009 | 21,405 | **21,405** |
+
+(All five 2026-09 revisions confirmed by re-running `prod_lines.py` against
+`git archive ee26268`; the pre-#785 column is a snapshot at an earlier commit,
+not a stale guess — it's what the tree actually was before #785 moved
+`struct App`.)
 
 **Projected vs. actual, over the chain (08-31 → 09-03):**
 
@@ -115,18 +122,28 @@ shared engine grew *more* than projected. Where the reduction actually came from
 `gtk/mod.rs` alone is 78% of the cut. Notably `tui_main/mouse.rs` — the file #733
 was sized against at −3,000…−3,500 — lost **316 lines**.
 
+**Attribution correction (#827): the −3,656 is mostly dead-code removal, not
+convergence.** Of the −3,656 backend reduction, roughly **−2,825** is #731
+(orphan Relm4 handles) and #732 (the GTK `Msg` bus) deleting code outright —
+that is almost all of the −2,834 booked against `src/gtk/mod.rs` above.
+Convergence proper (#751–#766 actually sharing logic through `render.rs`)
+moved something closer to **900 lines** out of the backends, while
+`render.rs` absorbed roughly **7,000**. State it plainly: the chain did not
+converge anywhere near as much as the raw −3,656/+6,396 pair implies: most of
+the shrinkage is code that was simply dead.
+
 > **#785 (stage 1 of #47) moved the mass, it did not delete it.** `struct App`,
 > its `impl` blocks and `impl quadraui::ShellApp for App` were hoisted verbatim
 > out of `src/gtk/mod.rs` into a new top-level `src/app.rs`. `src/gtk/`
-> therefore reads **2,584** against the 9,650 column above and `src/app.rs`
-> reads 7,131 — but the *total* is unchanged (+65, all of it module doc and
-> re-stated imports). Nothing here got smaller; a ~6,900-line block that was
+> reads **2,607** against the pre-#785 9,650 column above and `src/app.rs`
+> reads **7,131** — both regenerated at `ee26268` in the table above — but the
+> *total* is essentially unchanged (module doc and re-stated imports account
+> for the difference). Nothing here got smaller; a ~6,900-line block that was
 > filed under "GTK backend" is now filed under "shell application", where a
 > second native backend can reach it. `src/app.rs` is still `gui`-gated: its
 > module doc enumerates the four platform-typed fields, ~11 platform hook call
 > sites and the `crate::gtk::{click, css, util}` dependency that have to go
-> before the gate can. Read the backend columns above as *pre-#785* until they
-> are regenerated.
+> before the gate can.
 
 > **Correcting the record.** The figure this file previously carried as
 > "`src/gtk/` = 12,588 at 2026-09-01" was measured *before* #727/#728/#730 landed;
@@ -152,13 +169,17 @@ the remaining gap is small.
 
 ### 1. The irreducible surface — ✅ aggregated, and it is small
 
-Done: **[`docs/IRREDUCIBLE_SURFACE.md`](docs/IRREDUCIBLE_SURFACE.md)** (2026-09-03).
+Done: **[`docs/IRREDUCIBLE_SURFACE.md`](docs/IRREDUCIBLE_SURFACE.md)** (2026-09-03,
+corrected 2026-09-05 per #827 — the folder-picker verdict below was wrong).
 The headline, because it changes how the rest of this goal should be planned:
 
-- The **nine** recorded verdicts reduce to **four distinct facts**, of which **three are
-  genuinely irreducible** — the TUI-only folder picker (GTK uses a native
-  `GtkFileChooser`), px-vs-cell frame metrics, and GTK's menu bar *being* its CSD
-  titlebar (#552).
+- The **nine** recorded verdicts reduce to **three distinct facts**, of which **two are
+  genuinely irreducible** — px-vs-cell frame metrics, and GTK's menu bar *being* its CSD
+  titlebar (#552). A fourth candidate, the TUI-only folder picker, was recorded as
+  irreducible on the theory that GTK's native `GtkFileChooser` has no shared
+  counterpart — wrong: `quadraui::compose::FolderPickerController` has existed since
+  2026-05-25 and its module doc explicitly tells vimcode to delete the local copy and
+  rewire both backends through it. That verdict is struck, not counted.
 - **Only 1.3% of the two backends names a native toolkit type** — 246 production lines
   out of 19,429 (`scripts/native_lines.py`; ~25% undercount on the GTK side for stored
   widget handles, so call it under 4% even pessimistically).
@@ -170,7 +191,7 @@ The headline, because it changes how the rest of this goal should be planned:
 **Plan accordingly: this is ordinary duplication, not a platform-porting problem.** A
 plan that sizes it as the latter will keep missing its projection the way #751–#766 did.
 
-**The fourth verdict was mislabelled** — `tui_main/mouse.rs:1620` (command-line text
+**The third verdict was mislabelled** — `tui_main/mouse.rs:1620` (command-line text
 selection) reads as a decision but its own text says the fix is a
 `CommandLineLayout::hit_test` in quadraui. That is a *blocked* convergence whose blocker
 was never filed; `CommandLineLayout` does not exist in quadraui and **#194** is the
@@ -179,43 +200,63 @@ milestone-discipline rule at the bottom of this file, which applies to in-code c
 too: **a comment naming a missing upstream API is an unfiled issue, and grep will not
 find it for you.**
 
-### 2. The duplication moved down a level, into quadraui — and it is unqueued
-
-The two epics that now hold the mass of the remaining cross-backend duplication are
-**open, un-milestoned, and in nobody's queue**:
-
-- **quadraui#481** — *"shared runtime core: one implementation of the duplicated 65%
-  across backends."* 1,671 non-trivial lines byte-identical between `gtk/*.rs` and
-  `macos/*.rs`; `EventOutcome` declared twice verbatim; the 120 ms resize debounce
-  written twice and absent on macOS; `shell_runner` 45 identical lines ×2. The
-  copies have already drifted.
-- **quadraui#482** — *"Backend API integrity: units, symmetry, error channel,
-  panic-free text paths."* Trait asymmetry, off-trait rasterisers, unit leaks, and a
-  UTF-8 boundary fix that exists as **7 private copies and zero public ones**.
+### 2. "The duplication moved down a level" is largely refuted (#827)
 
-This is the same supply-side trap as before, one level down: vimcode's backends got
-thinner by pushing logic into a crate whose *own* backends are duplicated.
+The previous revision of this file claimed quadraui#481/#482 held a large, unqueued
+mass of cross-backend duplication one level down. Re-checked against quadraui's own
+pinned rev (`42e0f8f`) on 2026-09-05: most of the individual claims do not hold up.
 
-### 3. #47 was closed without its blocker
-
-**#47 (native macOS GUI) is closed and shipped zero code.** Commit `44882e9`
-("re-audit at pickup, no code — Backend-trait Rc-handle gap blocks Stage 1")
-recorded the real blocker: `App` calls `GtkBackend::modal_stack_handle()` /
-`drag_state_handle()` at 44 call sites in the drag/modal dispatch paths. Those are
-**inherent methods on the concrete struct, not on `quadraui::Backend`**, and
-`MacBackend`'s trait equivalents return short-lived `&mut` borrows incompatible
-with `App`'s stash-then-reuse pattern. Full findings and two candidate API shapes
-are in `PLAN.md`.
-
-That commit's own recommendation — *"file this as a quadraui issue before any
-vimcode-side Stage 1 code is written"* — **was never carried out.** No open
-quadraui issue mentions `modal_stack_handle` or `drag_state_handle`. The finding
-now lives only in `PLAN.md`, attached to a **closed** issue, which is exactly where
-a later triage pass will not look.
-
-**Action:** file the gap on `JDonaghy/quadraui` (or fold it into quadraui#482, its
-natural home), re-open quadraui milestone **#9 "vimcode Platform-Neutral
-blockers"**, and re-open vimcode#47 behind it.
+| Claim (previous revision) | Reality at pin `42e0f8f` |
+|---|---|
+| `EventOutcome` declared twice verbatim | Declared **once** — `quadraui/src/runtime.rs:95` (quadraui#496, closed 09-02) |
+| `shell_runner` 45 identical lines ×2 | Four runners of 4–7 non-comment lines each, all delegating to shared `shell_adapter.rs::build_shell_adapter` |
+| 1,671 byte-identical lines gtk↔macos | Function-level duplication is ~85 lines. quadraui#481's own correction comment (09-03 17:45Z) withdrew the headline number as "idiom coincidence" |
+| UTF-8 fix: 7 private copies, 0 public | **Public since 2026-08-15** — `text_util.rs:51-107`, re-exported from `lib.rs` (quadraui#503) |
+| `gtk_tree_layout`/`mac_tree_layout` twins | Both 1-line wrappers over `primitives/layout_metrics.rs:60 tree_layout` (quadraui#499, 09-02) |
+| "no `desktop/`" | `quadraui/src/desktop.rs` exists (754 lines) since 09-02 (quadraui#498) — the original claim grepped for a directory that had been renamed |
+| #482 "holds the mass" | **All eight children #503–#510 are closed.** #482 has zero comments and is a hollow epic |
+
+**Still true:** macOS dispatches `WindowResized` undebounced (`macos/run.rs:544-561`)
+while TUI/GTK use the shared `ResizeDebouncer`. That is a real, small, still-open gap —
+just not the "65% duplicated" epic the previous revision described. quadraui#481/#482
+remain open and un-milestoned, but do not plan against their headline numbers; re-audit
+the specific claim you need before acting on it.
+
+### 3. #47's blocker was filed and cleared — this section was stale for two days (#827)
+
+**Corrected 2026-09-05.** The previous revision said *"No open quadraui issue
+mentions `modal_stack_handle` or `drag_state_handle`"* and called filing one "the
+single most actionable item on this page." That stopped being true within hours of
+being written:
+
+- **quadraui#699** was filed 2026-09-03 **16:38Z**, into quadraui milestone **#9**,
+  and **closed 17:11Z** (PR#700 / `88345fb`). A follow-up, **#704**, closed 21:41Z.
+- **vimcode#47 was reopened 16:38Z** and is **OPEN**, in milestone **#5**, right now
+  — it was never re-closed.
+- The commit that last touched this file (`5e2c7cc`) landed **18:32Z — 81 minutes
+  after #699 had already closed** — and still said the blocker was unfiled. The
+  finding sat in `PLAN.md` for less than a day before it was acted on; the doc that
+  said otherwise just never got re-read against events.
+- quadraui milestone #9 was never closed — it's **open** (0 open / 7 closed issues
+  in it). The previous instruction to "re-open" it was acting on a wrong premise.
+
+**What actually happened, and what's still open:** quadraui#699/#704 gave every
+backend a symmetric Rc-handle API (`modal_stack_handle()` / `drag_state_handle()`),
+removing the `Backend`-trait asymmetry that blocked Stage 1. vimcode has already
+started consuming it — **#811** (this branch's own history) bumped the quadraui pin
+to `4ff2a64` and ported the four TUI-side call sites off the now-removed
+`drag_and_modal_mut`. **vimcode#47 itself is still open** (Stage 1 — moving `App`'s
+remaining GTK-specific call sites onto the new API, see `PLAN.md`) and is the actual
+next actionable item here, not a re-filing task.
+
+**The "44 call sites" figure was also wrong**, independent of the above. It came
+from `grep -n 'self\.backend\.' src/gtk/mod.rs` — every use of the `backend` field,
+not just the two Rc-handle methods. The real count, measured at `ee26268`
+(`grep -n 'modal_stack_handle\|drag_state_handle' src/app.rs`, minus the two
+doc-comment mentions of the method names): **19** — `modal_stack_handle` ×12,
+`drag_state_handle` ×7. (Also note the field moved: by `ee26268` this code lives in
+`src/app.rs`, not `src/gtk/mod.rs` — #785 had already hoisted it. Fixed in `PLAN.md`
+and `PROJECT_STATE.md` too, which is where the 44 figure originates.)
 
 ### 4. The divergence bug class is not dead
 
@@ -236,8 +277,8 @@ count.
 |---|---|---|
 | **quadraui#465** | macOS `ShellApp` + `run_with_shell` composition | ✅ Closed 2026-08-31. The supply-side gate is cleared. |
 | **#657** | Put vimcode on the oracle loop | ✅ Closed. `[lib] vimcode_core` + sealed `tests/acceptance/`. |
-| **#47** | Native macOS GUI, as a thin wrapper | ⚠️ **Closed with no code.** Blocker unfiled — see above. |
-| **quadraui#481 / #482** | The remaining duplication, one level down | 🔓 Open, un-milestoned, unqueued. |
+| **#47** | Native macOS GUI, as a thin wrapper | 🔓 **Reopened 2026-09-03, OPEN in milestone #5.** Blocker resolved (quadraui#699/#704); #811 already ported the TUI side. Stage 1 (GTK side) is the actual next work — see `PLAN.md`. |
+| **quadraui#481 / #482** | Duplication one level down — largely refuted, see §2 above | 🔓 Open, un-milestoned. Don't plan against their headline numbers. |
 
 ### The two decisions this file was holding open — both now moot
 
@@ -252,29 +293,36 @@ every fix ahead of #657 was verified by tests its own author wrote. It is now
 follow-up is to decide whether any of #751–#766 warrants a retro-fitted
 oracle-authored test, rather than re-litigating the sequencing.
 
-## Status (2026-09-03)
+## Status (2026-09-05, corrected per #827)
 
 - ✅ **Milestone #7 is 0 open.** The 09-01 critical path plus 16 slices all landed.
 - ✅ **The oracle loop is live here** (#657) and `draw_frame` is gone (#766).
-- 📉 **The audit is run and it missed:** backends −3,656 against a −8,700…−9,500
-  projection; the three files net **+2,740** lines.
-- ⚠️ **#47 closed with zero code and its blocker filed nowhere** — the single most
-  actionable item on this page.
-- 🔓 **quadraui#481 / #482 hold the remaining duplication** and are unqueued.
-- ✅ **The irreducible surface is aggregated** — [`docs/IRREDUCIBLE_SURFACE.md`](docs/IRREDUCIBLE_SURFACE.md).
-  Only **1.3%** of the backends is platform-bound; the rest is duplication, and the goal
-  should be planned as such.
+- 📉 **The audit is run and it missed its projection** — backends −3,656 against a
+  −8,700…−9,500 projection; the three files net **+2,740** lines. But ~−2,825 of
+  that −3,656 is #731/#732 dead-code removal, not convergence — convergence proper
+  moved roughly **900 lines** out of the backends while `render.rs` absorbed **~7,000**.
+- 🔓 **#47 is open again, in milestone #5.** Its blocker (quadraui#699/#704) closed
+  2026-09-03; #811 already ported the TUI side onto the new API. Stage 1 (GTK side)
+  is the actual next actionable item — not a re-filing task.
+- 🔓 **quadraui#481 / #482 remain open and un-milestoned**, but most of their
+  headline duplication claims were refuted 2026-09-05 (see §2) — the one confirmed
+  live gap is macOS's undebounced `WindowResized`.
+- ✅ **The irreducible surface is aggregated** — [`docs/IRREDUCIBLE_SURFACE.md`](docs/IRREDUCIBLE_SURFACE.md),
+  corrected 2026-09-05 (the folder-picker verdict was wrong; struck). **Two** of the
+  three remaining facts are genuinely irreducible; only **1.3%** of the backends
+  names a native toolkit type — the rest is duplication, and the goal should be
+  planned as such.
 
 ## How to use this doc
 
 - **Line numbers:** this file cites none, on purpose — locate by symbol
-  (`grep -n "impl quadraui::ShellApp for App" src/gtk/mod.rs`). Counts here are
+  (`grep -n "impl quadraui::ShellApp for App" src/app.rs`). Counts here are
   evidence measured on a named revision; **regenerate them with
-  `python3 scripts/prod_lines.py src/gtk src/tui_main src/render.rs`** rather than
-  trusting the table.
+  `python3 scripts/prod_lines.py src/gtk src/tui_main src/render.rs src/app.rs`**
+  rather than trusting the table.
 - **Agents:** treat this as the standing objective behind all planning and triage.
-  There is no queue to work right now — the next move is item 3 above (file the
-  quadraui gap), then items 1 and 2. Never write new per-backend code
+  Item 3's blocker is cleared — the next move is **vimcode#47 Stage 1** (see
+  `PLAN.md`), then items 1, 2 and 4 above. Never write new per-backend code
   (`CLAUDE.md` Platform-Neutrality Rule). When you adopt a quadraui API,
   **delete** the old backend code in the same PR.
 - **Humans:** edit freely as priorities shift; keep it short, re-date Status.
diff --git a/PLAN.md b/PLAN.md
index 0e8fc8aa..b0f6d271 100644
--- a/PLAN.md
+++ b/PLAN.md
@@ -6,30 +6,43 @@
 > source of truth for individual tasks — this file points at the current
 > wave and explains how to resume.
 >
-> **Last updated:** 2026-09-03 — **#47 re-audited, never started, and now
-> CLOSED with no code.** The chain (#730→#657) plus slices #751–#766 are fully
-> landed and milestone #7 is 0 open, but the re-audit #47 mandates at pickup
-> found the real blocker is a `Backend`-trait ergonomics mismatch the issue text
-> didn't anticipate. **The quadraui issue that re-audit told us to file was never
-> filed, and #47 was closed anyway** — so this section is the only surviving
-> record of the blocker. Read it before attempting Stage 1, and file the gap
-> upstream first.
+> **Last updated:** 2026-09-05 (issue #827 correction pass). **The blocker below
+> was filed and cleared 2026-09-03 — this file just never caught up.** The
+> 2026-09-03 revision said "the quadraui issue was never filed" and counted "44"
+> `self.backend` call sites; both were wrong (the real blocker, quadraui#699, was
+> filed and closed the same day at 16:38Z–17:11Z, and the real count of the two
+> Rc-handle methods is **19**, not 44 — see the corrected blocker section below).
+> **#47 is reopened, in milestone #5**, and #811 (already on this branch) has
+> started consuming the fix on the TUI side. Stage 1 (the GTK-side `App` move)
+> is the live next step — read the section below for what's still accurate.
 
 ---
 
-## ⚠️ Live blocker — **#47 (macOS wrapper): closed with no code, blocker unfiled**
-
-**Read this first.** #47 is *closed*, but nothing was built and the design question
-below is still open. No open quadraui issue mentions `modal_stack_handle` or
-`drag_state_handle`; `git show 44882e9` is a documentation-only diff. Before any
-macOS work starts:
-
-1. File the `Backend`/`GtkBackend` symmetry gap on `JDonaghy/quadraui` (or fold it
-   into **quadraui#482**, "Backend API integrity", its natural home).
-2. Re-open quadraui milestone **#9 "vimcode Platform-Neutral blockers"** to hold it.
-3. Re-open **vimcode#47** behind that blocker rather than leaving it closed —
-   `GOALS.md` now states the rule this violated: *a #7 issue that turns out to be
-   supply-blocked stays open behind its blocker.*
+## ✅ Blocker resolved 2026-09-03 — **#47 (macOS wrapper) is reopened, in milestone #5**
+
+**Corrected 2026-09-05 (issue #827).** This section previously said #47 was closed
+with its blocker unfiled. Both halves of that were wrong:
+
+- **quadraui#699** — the `Backend`/`GtkBackend` symmetry gap below — was filed
+  2026-09-03 **16:38Z** (into quadraui milestone #9) and **closed 17:11Z**
+  (PR#700/`88345fb`). Follow-up **#704** closed 21:41Z.
+- **vimcode#47 was reopened 16:38Z** and is open now, in milestone #5 — it was
+  never left closed behind an unfiled blocker.
+- quadraui#699/#704 gave every backend a symmetric Rc-handle API
+  (`modal_stack_handle()` / `drag_state_handle()`), which is what the design
+  discussion below asked for (option 1). **#811** (already on this branch) has
+  ported the TUI side of vimcode onto it — bumped the quadraui pin to `4ff2a64`
+  and rewrote the four `drag_and_modal_mut` call sites in
+  `src/tui_main/shell_app.rs`.
+
+**What's still open:** the rest of Stage 1 — `struct App` (`modal_stack_handle()`
+×12 + `drag_state_handle()` ×7 = 19 call sites, not 44 — see below) already moved
+into `src/app.rs` via #785, and #811 already ported the *TUI*-relevant call sites;
+what's left is cfg-gating `App`'s remaining GTK-only fields and finishing the
+`App`-move plan below. The design analysis below (why `modal_stack_mut`/
+`drag_and_modal_mut` couldn't be stashed) is preserved as background for *why* the
+fix took the shape it did; treat "no code was moved" and "the upstream issue was
+never filed" (further down) as no longer true.
 
 _As of 2026-09-02._ Picked up #47 (native macOS GUI, thin wrapper over
 `quadraui::macos::shell_runner::run_with_shell`). The issue's own body flags
@@ -69,7 +82,11 @@ verbatim; that part of Stage 1 is low-risk mechanical work.
 
 **The actual blocker is `backend`.** `backend::GtkBackend` is a re-export of
 `quadraui::gtk::GtkBackend` (`src/gtk/backend.rs`). `App` calls exactly five
-methods on it (`grep -n 'self\.backend\.' src/gtk/mod.rs`, 44 call sites):
+methods on it (`grep -n 'self\.backend\.' src/gtk/mod.rs`, **44 call sites** —
+**corrected 2026-09-05: that count is every use of the `backend` field across
+all five methods, not just the two Rc-handle ones this blocker is about. The
+count that actually matters is `modal_stack_handle` ×12 + `drag_state_handle`
+×7 = 19 call sites**, verified at `ee26268`):
 `modal_stack_handle()`, `drag_state_handle()`, `set_current_line_height()`,
 `set_current_char_width()`, `set_pango_context()` (the last is GTK/Pango-only,
 one call site at `:5548`, trivially cfg-gated). The first four are **inherent
@@ -101,9 +118,9 @@ to the `&mut dyn Backend` borrow's lifetime — it cannot be stashed and reused
 across the borrow-drop points this pattern relies on. Simply changing
 `App.backend`'s type from `Rc<RefCell<backend::GtkBackend>>` to
 `Rc<RefCell<Box<dyn quadraui::Backend>>>` (the obvious first move) does not
-compile: the trait object exposes the wrong shape for ~40 of the 44 call
-sites, and this is the *safety-critical* code (drag/modal/click dispatch),
-not a corner where a quick workaround is low-risk.
+compile: the trait object exposes the wrong shape for the 19 `modal_stack_handle`/
+`drag_state_handle` call sites, and this is the *safety-critical* code
+(drag/modal/click dispatch), not a corner where a quick workaround is low-risk.
 
 **This is a supply-side gap, not a vimcode design choice** — the
 Platform-Neutrality Rule says stop and file it upstream rather than route
@@ -120,33 +137,40 @@ should pick, not a vimcode-side workaround):
    needs to read them for hit-testing) — bigger change, but removes the
    asymmetry at the root instead of papering over it per-backend.
 
-**Recommendation:** file this as a quadraui issue (gap in `Backend`/`GtkBackend`
-API symmetry) before any vimcode-side Stage 1 code is written, per the
-Platform-Neutrality Rule. Do **not** attempt the `App` move without that
-decision — the 44 `self.backend` call sites sit in the app's most
-safety-critical dispatch paths (mouse drag, modal overlays), and guessing at
-an API shape here risks the same "three attempts, three reverts" outcome
-`CLAUDE.md`'s #319 negative example documents for exactly this kind of
-per-backend improvisation.
-
-**What did NOT happen this session — and has not happened since:** no code was
-moved, and the upstream issue was never filed. `struct App` and
-`impl quadraui::ShellApp for App` are untouched in `src/gtk/mod.rs`. The three
+**Recommendation (as of the 2026-09-02 re-audit):** file this as a quadraui issue
+(gap in `Backend`/`GtkBackend` API symmetry) before any vimcode-side Stage 1 code
+is written, per the Platform-Neutrality Rule. That happened — see the corrected
+"Blocker resolved" section at the top of this file: **quadraui#699/#704 shipped
+the Rc-handle accessors on option 1's shape**, and #811 already ported the TUI
+side onto them. The 19 `modal_stack_handle`/`drag_state_handle` call sites in
+`src/app.rs` (moved there by #785 since this analysis was written) are what
+Stage 1's GTK-side port still needs to touch.
+
+**Status, corrected 2026-09-05:** the upstream issue was **not** left unfiled —
+quadraui#699 was filed and closed the same day (2026-09-03, 16:38Z–17:11Z). `struct
+App` and `impl quadraui::ShellApp for App` moved (verbatim, per #785) out of
+`src/gtk/mod.rs` into `src/app.rs` before this correction was written, and #811 has
+already rewritten the TUI-relevant call sites onto the new handle API. The three
 other platform fields (`window`, `css_provider`, `settings_monitor`) and the
-key-mapping helpers are confirmed low-risk and ready to move once the
-`backend` question is resolved — re-verify line numbers again at that point,
-since nothing here has landed to keep them stable.
-
-### Stage ordering once unblocked
-
-**Step 1 has still not happened** (verified 2026-09-03: no open quadraui issue
-references either accessor). Nothing below is startable until it does.
-
-1. File/resolve the quadraui `Backend`/`GtkBackend` symmetry gap above.
-2. Move `struct App` + its three `impl App` blocks + `impl ShellApp for App`
-   (`src/gtk/mod.rs:538-7483`) to `src/app.rs`, gating `window`/`css_provider`/
-   `settings_monitor` behind `#[cfg(feature = "gui")]` and switching `backend`
-   to whatever shape step 1 lands on.
+key-mapping helpers are still confirmed low-risk and ready to move — re-verify line
+numbers in `src/app.rs` before touching them, since #785/#811 already changed them
+once since this analysis was written.
+
+### Stage ordering — unblocked 2026-09-03
+
+**Step 1 is resolved** (quadraui#699/#704 closed 2026-09-03; #811 ported the TUI
+side). Re-audit which of the steps below `#785`/`#811` already completed before
+starting the next one — this list was written before either landed.
+
+1. ~~File/resolve the quadraui `Backend`/`GtkBackend` symmetry gap above.~~ ✅ Done
+   (quadraui#699/#704).
+2. ~~Move `struct App` + its three `impl App` blocks + `impl ShellApp for App`
+   (`src/gtk/mod.rs:538-7483`) to `src/app.rs`~~ ✅ **The move itself is done**
+   (#785, verbatim). **Still open:** gating `window`/`css_provider`/
+   `settings_monitor` behind `#[cfg(feature = "gui")]` and switching `backend`'s
+   ~19 `modal_stack_handle`/`drag_state_handle` call sites onto the new handle
+   API from quadraui#699/#704 (per `GOALS.md`, `src/app.rs`'s own module doc
+   still lists these as the remaining gate-blockers).
 3. `src/gtk/mod.rs` keeps only `run()`, `build_shell_config()`, and the
    genuinely GTK-only helpers (window-chrome, CSS, key-name→GDK glyph tables
    if any remain GTK-specific after the move).
diff --git a/PROJECT_STATE.md b/PROJECT_STATE.md
index d823cb9b..8e30470e 100644
--- a/PROJECT_STATE.md
+++ b/PROJECT_STATE.md
@@ -1,6 +1,6 @@
 # VimCode Project State
 
-**Last updated:** September 4, 2026 (see Recent Work for #801); prior revision September 3, 2026 — **the platform-neutrality chain drained, and the audit it mandated is now run.** Milestone #7 is **0 open**: everything the 2026-09-01 audit filed landed, including 16 slices it never named (#751–#766). The post-#735 sizing audit — which the previous revision explicitly warned not to skip — is below, and it **missed its projection by roughly 60%**. Nothing is in flight and nothing is queued for vimcode or quadraui. The most actionable thing on this page is that **#47 was closed having shipped zero code, with its blocker filed nowhere.**
+**Last updated:** September 5, 2026 (issue #827 correction pass — the #47/44-call-site claims below were stale within hours of being written; see the corrected section); prior revisions September 4 (#801) and September 3 (the platform-neutrality chain drained, and the audit it mandated is now run). Milestone #7 is **0 open**: everything the 2026-09-01 audit filed landed, including 16 slices it never named (#751–#766). The post-#735 sizing audit — which the previous revision explicitly warned not to skip — is below, and it **missed its projection by roughly 60%** (though most of that miss is #731/#732 dead-code removal, not convergence — see `GOALS.md` §2/§3 for the corrected attribution). Nothing is in flight and nothing is queued for vimcode. **#47 is reopened, in milestone #5** — its blocker (quadraui#699/#704) was filed and closed 2026-09-03, and #811 already ported the TUI side onto the new API. See `GOALS.md` for the full correction.
 
 ## Active milestone: #7 Platform-Neutral — **complete (0 open)**
 
@@ -44,15 +44,21 @@ Still true from earlier in the arc: `fn event_loop` does not exist in `src/`;
 
 ### The post-#735 sizing audit — run on `develop @ eedebf8`
 
-Production lines, `#[cfg(test)]` excluded. **All four columns measured with the same
+Production lines, `#[cfg(test)]` excluded. **All columns measured with the same
 script** (`scripts/prod_lines.py`, added for this audit) so they are comparable:
 
-| | 2026-05-01 | 2026-07-01 | pre-chain 2026-08-31 | **now 2026-09-03** |
-|---|---|---|---|---|
-| `src/gtk/` | 18,969 | 13,675 | 12,526 | **9,650** |
-| `src/tui_main/` | 14,649 | 10,358 | 11,125 | **10,345** |
-| **both backends** | 33,618 | 24,033 | 23,651 | **19,995** |
-| `src/render.rs` (shared) | 10,574 | 12,807 | 15,009 | **21,405** |
+| | 2026-05-01 | 2026-07-01 | pre-chain 2026-08-31 | pre-#785 2026-09-03 | **post-#785, now @ `ee26268`** |
+|---|---|---|---|---|---|
+| `src/gtk/` | 18,969 | 13,675 | 12,526 | 9,650 | **2,607** |
+| `src/tui_main/` | 14,649 | 10,358 | 11,125 | 10,345 | **10,366** |
+| `src/app.rs` (hoisted out of `src/gtk/` by #785) | — | — | — | — | **7,131** |
+| **all three files** | 33,618 | 24,033 | 23,651 | 19,995 (2 files) | **20,104** |
+| `src/render.rs` (shared) | 10,574 | 12,807 | 15,009 | 21,405 | **21,405** |
+
+(#785, "stage 1 of #47," hoisted `struct App` verbatim out of `src/gtk/mod.rs` into
+a new `src/app.rs` — see `GOALS.md`'s post-#735 audit for the full account. The
+`src/gtk/` = 9,650 figure this file previously carried as "now" predates that move;
+regenerated at `ee26268` per #827.)
 
 **Projected vs. actual over the chain (08-31 → 09-03):**
 
@@ -98,8 +104,11 @@ is a long way from the north star, and the remaining gap should not be planned a
 ### What remains — four items, none of them queued
 
 1. ~~**The irreducible surface is recorded but never aggregated.**~~ ✅ **Done
-   2026-09-03: [`docs/IRREDUCIBLE_SURFACE.md`](docs/IRREDUCIBLE_SURFACE.md).** The nine
-   verdicts reduce to four facts, three genuinely irreducible. And the sizing answer:
+   2026-09-03, corrected 2026-09-05:
+   [`docs/IRREDUCIBLE_SURFACE.md`](docs/IRREDUCIBLE_SURFACE.md).** The nine
+   verdicts reduce to **three** facts (the folder-picker verdict was wrong and has
+   been struck — `quadraui::compose::FolderPickerController` has existed since
+   2026-05-25), **two** genuinely irreducible. And the sizing answer:
    **only 246 of 19,429 production lines (1.3%) name a native toolkit type**, so
    platform-specificity is *not* what keeps the backends large — `src/gtk/mod.rs` and
    `src/tui_main/shell_app.rs` are two implementations of the same four `ShellApp` entry
@@ -107,52 +116,55 @@ is a long way from the north star, and the remaining gap should not be planned a
    (`tui_main/mouse.rs:1620`, command-line selection) turned out to be a **mislabelled
    supply gap**: `CommandLineLayout::hit_test` does not exist in quadraui and was never
    filed; **#194** is the open consumer-side symptom.
-2. **The duplication moved down into quadraui and is unqueued.** **quadraui#481**
-   (shared runtime core — 1,671 non-trivial lines byte-identical between `gtk/*.rs` and
-   `macos/*.rs`, `EventOutcome` declared twice verbatim, the resize debounce written
-   twice and absent on macOS) and **quadraui#482** (Backend API integrity — trait
-   asymmetry, unit leaks, a UTF-8 boundary fix living as 7 private copies). Both open,
-   both un-milestoned, neither in any queue.
-3. **#47's blocker is filed nowhere** — see below.
+2. **The "duplication moved down into quadraui" claim is largely refuted (#827).**
+   quadraui#481/#482 remain open and un-milestoned, but most of the headline numbers
+   don't hold up at the pinned rev: `EventOutcome` is declared once, not twice
+   (quadraui#496); the 1,671-line byte-identical claim was withdrawn by quadraui#481's
+   own correction comment as "idiom coincidence" (real duplication ~85 lines); the
+   UTF-8 fix has been public since 2026-08-15 (quadraui#503); the tree-layout
+   "twins" are both 1-line wrappers over one shared function (quadraui#499); and
+   quadraui#482's eight children (#503–#510) are all closed. See `GOALS.md` §2 for
+   the full table. What's still real: macOS dispatches `WindowResized` undebounced
+   while TUI/GTK share a `ResizeDebouncer`.
+3. **#47's blocker was filed and cleared 2026-09-03** — see below (this used to say
+   "filed nowhere"; it wasn't, within hours of that claim being written).
 4. **The divergence bug class is still ~44 issues deep** (#206, #420, #264, #194, #233
    and friends), plus milestone #5's cross-backend residue (#149, #167, #168, #233,
    #294). `GOALS.md`'s thesis is that each is a symptom of a duplicated surface; if the
    convergence had reached far enough this list would be shrinking. It is the only
    outcome measure this goal has that isn't a line count — watch it.
 
-### ⚠️ #47 was closed having shipped no code
+### ✅ #47's blocker was filed and cleared — this section was stale (#827)
 
-**#47 (native macOS GUI) is closed and its diff is documentation only.** Commit
-`44882e9` — *"re-audit at pickup, no code — Backend-trait Rc-handle gap blocks Stage 1"* —
-recorded the real blocker: `App` calls `GtkBackend::modal_stack_handle()` /
-`drag_state_handle()` at 44 call sites in the drag and modal dispatch paths. Those are
+**Corrected 2026-09-05.** #47 (native macOS GUI) was closed 2026-09-02 with commit
+`44882e9` — *"re-audit at pickup, no code — Backend-trait Rc-handle gap blocks Stage
+1"* — recording the real blocker: `App` called `GtkBackend::modal_stack_handle()` /
+`drag_state_handle()` at **19** call sites (`modal_stack_handle` ×12,
+`drag_state_handle` ×7 — not the "44" this file previously said, which counted every
+use of the `backend` field via `grep -n 'self\.backend\.' src/gtk/mod.rs`, not just
+the two Rc-handle methods) in the drag and modal dispatch paths. Those were
 **inherent methods on the concrete struct, not on the generic `quadraui::Backend`
 trait**, and `MacBackend`'s trait equivalents (`modal_stack_mut`, `drag_and_modal_mut`)
-return short-lived `&mut` borrows that cannot be stashed and reused the way `App` does.
-Full findings and two candidate API shapes are in [`PLAN.md`](PLAN.md).
-
-That commit's own recommendation — *"file this as a quadraui issue before any
-vimcode-side Stage 1 code is written"* — **was never carried out.** No open quadraui
-issue mentions either method name. The finding now lives only in `PLAN.md`, attached to a
-**closed** issue, which is precisely where the next triage pass will not look.
-
-**This is `GOALS.md`'s supply-side trap in a new shape.** The documented failure mode was
-"infra lands in quadraui but the #7 adoption issue never gets picked up." This is the
-inverse: the consume-side issue was *closed* while its supply-side blocker went
-unrecorded. The rule that would have caught it is now written down in `GOALS.md`: **a #7
-issue that turns out to be supply-blocked stays open behind its blocker; it does not get
-closed.**
-
-**Action:** file the gap on `JDonaghy/quadraui` (or fold it into quadraui#482, its
-natural home), re-open quadraui milestone #9 "vimcode Platform-Neutral blockers", and
-re-open vimcode#47 behind it.
+returned short-lived `&mut` borrows that couldn't be stashed and reused the way `App`
+does. Full findings are in [`PLAN.md`](PLAN.md).
+
+**That blocker was filed — this file just never caught up.** **quadraui#699** was
+filed 2026-09-03 16:38Z (into quadraui milestone #9) and **closed 17:11Z**
+(PR#700/`88345fb`); follow-up **#704** closed 21:41Z. **vimcode#47 was reopened
+16:38Z** and is **open now, in milestone #5**. quadraui#699/#704 gave every backend a
+symmetric Rc-handle API, and vimcode has already started consuming it: **#811**
+bumped the quadraui pin to `4ff2a64` and ported the TUI-side call sites off the
+now-removed `drag_and_modal_mut`. The actual next actionable item is **vimcode#47
+Stage 1** (the GTK-side `App` move), not a re-filing task — see `PLAN.md` and
+`GOALS.md` for the full correction.
 
 ### Milestone hygiene
 
 - **#7 is 0 open.** #146 moved to #4 Editor Features; #47 sits in #5 Cross-Platform UI
-  Crate, which `GOALS.md` defines as covering the macOS/Windows backends.
-- **quadraui milestone #9** ("vimcode Platform-Neutral blockers") is closed out and
-  should be re-opened to hold the #47 blocker.
+  Crate, which `GOALS.md` defines as covering the macOS/Windows backends, and is
+  **open** (reopened 2026-09-03).
+- **quadraui milestone #9** ("vimcode Platform-Neutral blockers") is **open** (0
+  open / 7 closed issues) — it held quadraui#699 and does not need re-opening.
 - **Stale Win-GUI issues.** Roughly a dozen open `Win-GUI:` issues (#160–#178, #61,
   #172, #176) describe a backend **deleted from this repo on 2026-05-11** (`3e4bcff`).
   Their live counterparts are quadraui#19–#31 / quadraui#580. They should be migrated or
@@ -317,7 +329,20 @@ mouse routing, #757–#762 keyboard dispatch, #763–#766 frame composition (`Fr
 post-#735 sizing audit the previous revision mandated and added `scripts/prod_lines.py`
 so it is reproducible: backends **−3,656** against a −8,700…−9,500 projection, `render.rs`
 **+6,396**, net **+2,740**. #47 closed having shipped **no code** (`44882e9`) with its
-`Backend`-trait Rc-handle blocker filed nowhere — the top open action.
+`Backend`-trait Rc-handle blocker filed nowhere — the top open action. *(Corrected
+2026-09-05, issue #827: that blocker — quadraui#699, at 19 not 44 call sites — was
+filed and closed the same day, 16:38Z–17:11Z, and #47 was reopened 16:38Z. This
+entry's "filed nowhere" was already wrong by the time the revision carrying it was
+written; see the corrected section above.)*
+
+**2026-09-05 — GOALS.md/PLAN.md/PROJECT_STATE.md/IRREDUCIBLE_SURFACE.md corrected
+(#827).** A four-agent audit of `develop @ ee26268` found the planning docs
+materially stale: the #47-blocker-unfiled claim (quadraui#699 had already closed),
+the 44-call-site figure (real count 19), the quadraui#481/#482 "duplication moved
+down a level" claims (mostly refuted at the pinned rev), the `src/gtk/` size-table
+column (predated #785's move), and the `IRREDUCIBLE_SURFACE.md` folder-picker
+verdict (wrong — `FolderPickerController` has existed in quadraui since 05-25).
+Corrected all four docs; no code changed.
 
 **2026-09-01 — platform-neutrality audit, and everything it found is now queued.**
 Filed #730 (`ai_panel`), #731 (orphan handles), #732 (`Msg` bus), #733 (mouse routers),
diff --git a/docs/IRREDUCIBLE_SURFACE.md b/docs/IRREDUCIBLE_SURFACE.md
index 49f3f040..78729fe4 100644
--- a/docs/IRREDUCIBLE_SURFACE.md
+++ b/docs/IRREDUCIBLE_SURFACE.md
@@ -6,27 +6,42 @@
 > surface that **stays** — which is what would let anyone judge how far 19,995 production
 > lines is from the north star. This is that statement.
 >
-> _Measured 2026-09-03 on `develop` @ `8e333a8`. Regenerate, don't trust:
+> _Measured 2026-09-03 on `develop` @ `8e333a8`. Corrected 2026-09-05 (issue #827)
+> — the folder-picker row below was wrong; struck, not just re-verdicted. Regenerate,
+> don't trust:
 > `python3 scripts/prod_lines.py src/gtk src/tui_main src/render.rs` and
 > `python3 scripts/native_lines.py gtk src/gtk/*.rs`._
 
-## 1. The nine recorded verdicts are four facts
+## 1. The nine recorded verdicts are three facts
 
 ```
 grep -rn -iE "do not converge|not converged|one-sided|intrinsic difference" src/
 ```
 
 returns nine anchors, but several are cross-references to the same decision. They reduce
-to four:
+to three:
 
 | Fact | Anchors | Verdict |
 |---|---|---|
-| **The folder / workspace picker is TUI-only.** GTK opens a *native* `GtkFileChooser`, deferred through `PendingFileDialog` and run from `tick()`, so there is no GTK canvas surface to paint, hit-test or arbitrate. | 5 — `render.rs:2919` (keyboard), `render.rs:7242` (`FrameOp::FolderPicker`), `render.rs:21569` (mouse-arbitration test), `gtk/mod.rs:6498` (the no-op compose arm), `tui_main/mouse.rs:691` (input) | ✅ **Irreducible.** Converging it means *building* a GTK canvas picker — adding per-backend code to make two tables match. |
-| **Frame metrics are px on GTK, cells on TUI.** `FrameMetrics` carries only `line_height`/`char_width` and answers one question: "is this reserved band at least one line tall". Rect math stays per backend because Cairo painter-order and ratatui cell coalescence differ intrinsically. | 1 — `render.rs:7162` | ✅ **Irreducible.** Already reduced to the minimum: a unit, not a geometry. |
+| **Frame metrics are px on GTK, cells on TUI.** `FrameMetrics` carries only `line_height`/`char_width` and answers one question: "is this reserved band at least one line tall". Rect math stays per backend because Cairo painter-order and ratatui cell coalescence differ intrinsically. | 1 — `render.rs:7162` | ✅ **Irreducible.** Already reduced to the minimum: a unit, not a geometry. See §2b for a caveat on how thin that "unit" actually stays. |
 | **GTK's menu bar *is* its client-side titlebar.** `App::setup` pins `engine.menu_bar_visible = true` unconditionally (#552); TUI shows its menu row only in vscode-mode or via Alt. | 2 — `gtk/testing.rs:5944`, `tui_main/shell_app.rs:6032` | ✅ **Irreducible.** A property of CSD, not a transcription. Handled by fixture, not by branching production code. |
-| **Command-line text selection is TUI-only.** GTK has no `cmd_sel`/`cmd_dragging` state, no inverted-cell read-back, and paints its command line through `Surface::CommandLine`, which exposes no character-offset hit test. | 1 — `tui_main/mouse.rs:1620` | ❌ **Not irreducible — mislabelled.** See §2. |
+| **Command-line text selection is TUI-only.** GTK has no `cmd_sel`/`cmd_dragging` state, no inverted-cell read-back, and paints its command line through `Surface::CommandLine`, which exposes no character-offset hit test. | 1 — `tui_main/mouse.rs:1620` | ❌ **Not irreducible — mislabelled.** See §2a. |
 
-## 2. One of the four is a supply gap wearing a verdict's clothes
+### 1b. Struck: "the folder / workspace picker is TUI-only" (#827)
+
+A fourth row previously sat in the table above, verdicted ✅ **Irreducible**: *"GTK
+opens a native `GtkFileChooser`, deferred through `PendingFileDialog` and run from
+`tick()`, so there is no GTK canvas surface to paint, hit-test or arbitrate."* That
+verdict was wrong, not just stale — it never checked upstream. `quadraui::compose::FolderPickerController`
+has existed since **2026-05-25**, and its own module doc explicitly instructs
+consumers (vimcode named) to delete the local picker and rewire both backends through
+the shared controller — the aggregation that produced this row grepped `src/` for
+"do not converge" anchors and never looked at what quadraui already shipped. This is
+an ordinary **#7 Platform-Neutral adoption gap**, not an irreducible surface: file the
+adoption issue against `FolderPickerController`, don't carry this row forward as a
+settled fact.
+
+## 2a. One of the three is a supply gap wearing a verdict's clothes
 
 `tui_main/mouse.rs:1620` files itself under the same "recorded rather than converged"
 heading as the others, but its own text says the opposite:
@@ -48,6 +63,36 @@ supply-blocked stays open behind its blocker* — applies to in-code verdicts to
 comment that says "this needs a quadraui API" is an unfiled issue, and grep will not
 find it for you.**
 
+## 2b. The `unit_w`/`unit_h` seam is thinner in practice than §1 blesses it (#827)
+
+§1 treats `FrameMetrics`'s `unit_w`/`unit_h` convention as *the* unit seam that keeps
+frame-metrics irreducibility down to "a unit, not a geometry." That's the intent, but
+`render.rs` doesn't hold the line as cleanly as the fact-table implies. As measured
+2026-09-05:
+
+- The `unit_w: f32, unit_h: f32` parameter pair appears in **7 painter function
+  signatures** in `render.rs` (`hover_popup_to_quadraui_tooltip`,
+  `editor_hover_popup_paint`, `signature_help_to_quadraui_tooltip`,
+  `panel_hover_popup_paint`, `diff_peek_to_quadraui_tooltip`,
+  `draw_ai_sidebar_panel`, `tab_hover_tooltip_paint`), across roughly **33 uses** of
+  the two identifiers in expressions.
+- `render.rs` additionally carries **two explicit `char_width > 1.0` backend
+  sniffs** — `minimap_reserved_width` (`render.rs:10426`) branches min/max width in
+  pixels vs. columns on it, and the editor-viewport scrollbar reservation
+  (`render.rs:17256`) reserves 8px only "in the GTK backend (`char_width > 1.0`)".
+  Both are comments-as-documentation admitting the unit is not opaque to the
+  caller — code downstream of the seam still asks "am I GTK?" by proxy.
+- The same file carries **five paired per-backend policy tables** alongside the
+  unit convention — constants/branches that assume one shape for `char_width == 1.0`
+  (TUI) and another for `char_width > 1.0` (GTK), rather than deriving the answer
+  purely from the unit value.
+
+**Record the real state, not the aspiration:** `unit_w`/`unit_h` is a real and useful
+convention, but it is a *convention observed by callers*, not an enforced boundary —
+`render.rs` still contains explicit backend-identity branches hiding behind the unit
+parameter's name. Treat the frame-metrics row in §1 as "irreducible, and mostly but
+not entirely behind one seam."
+
 ## 3. How much of the backends is actually platform-bound
 
 Production lines that name a toolkit module or type (`gtk4::`/`gio::`/`glib::`/`gdk::`/
@@ -81,8 +126,10 @@ tripled, the figure stays under 4%.
 keeping it there.**
 
 The two backends are, to within about 1.3%, ordinary toolkit-free Rust that happens to
-live in a backend directory. The genuinely irreducible surface is the three facts in §1
-plus roughly 250 lines of native calls — a few hundred lines, not tens of thousands.
+live in a backend directory. The genuinely irreducible surface is the two irreducible
+facts in §1 (frame metrics, GTK's CSD menu bar) plus roughly 250 lines of native calls
+— a few hundred lines, not tens of thousands. (The folder picker is not a third: §1b
+struck it as an unactioned adoption gap, not an irreducible fact.)
 
 **What this does NOT claim:** that the remaining ~19,200 lines are mechanically
 convergeable. Some of it is real per-backend *structure* — GTK and TUI compose a frame
@@ -108,6 +155,10 @@ the actual remaining work.
    platform-porting problem, and a plan that treats them as one will keep missing its
    projection the way the #751–#766 chain did (−3,656 against −8,700…−9,500). Size it as
    what it is: two implementations of four entry points.
+4. **File the folder-picker #7 adoption issue.** `quadraui::compose::FolderPickerController`
+   has shipped since 2026-05-25 and nothing has adopted it — file the vimcode-side
+   deletion/rewire against it (§1b), rather than continuing to treat the TUI-only picker
+   as a settled irreducible fact.
 
 ## 6. Regenerating this