docs(plan): correct three claims before merge

- A4: branch protection IS applied (six required checks, strict, no force
  pushes); what is missing is enforce_admins, which is why the owner's
  direct pushes land. B0 #3 becomes the owner's decision on that.
- A5 / B0 #1 / CLAUDE.md: the owner has run the v0.3.3 installer, and
  installer/ is identical to v0.3.4 apart from the version number;
  platform and log still to be recorded for 11.9.
- D1: Reddit is unreachable from the cloud environment only; the owner's
  machine reads the thread's Atom feed. D2: the watch drafts replies for
  the owner instead of posting them.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

Author: superwilso

CLAUDE.md

@@ -17,7 +17,7 @@
 > | Offline gates | **549 Rust tests**, **41 harness scenarios**, 23 C/C++ files syntax-clean, **234 golden pixel hashes** — all green, all re-run 2026-09-14 |
 > | The headline feature | **USB-DAC → LDAC ran end to end** (2026-09-12, owner-reported, no log captured) |
 > | Landed since this banner | FM tuner, Bluetooth, NFC, playlists, on-screen keyboard, liked-songs sync, device settings, palettes (device-verified), type scale, **lyrics and search** |
-> | Still device-gated | [`docs/DEVICE_CHECKLIST.md`](docs/DEVICE_CHECKLIST.md) — and **11.9, the Windows installer, has never been executed once** while being the only Windows install path |
+> | Still device-gated | [`docs/DEVICE_CHECKLIST.md`](docs/DEVICE_CHECKLIST.md) — **11.9, the Windows installer** (the only Windows install path): the owner reports running the v0.3.3 installer, whose `installer/` is identical to v0.3.4's; platform and log not yet recorded |
 >
 > **The live documents, in order of what you probably want:**
 > - **What to do next, and the skins design system:** [`docs/PLAN_2026-09-14.md`](docs/PLAN_2026-09-14.md) — current audit + ordered plan.

docs/PLAN_2026-09-14.md

@@ -76,15 +76,19 @@ directory. The argument for keeping them was reviewability — but **`golden.txt
 does the regression job**, and it does it better, because it fails CI instead of waiting to be
 looked at. Delete both, `.gitignore` them, publish the render as a CI artifact.
 
-### A4. Every commit goes straight to `main`
+### A4. Branch protection is on — and the owner is exempt from it
 
-`AUDIT_2026-09-01` §D5 called branch protection *"the most important item in this document"* and
-*"the one that cannot be fixed from a shell."* It was never recorded as applied. Measured today:
-**0 open PRs, and every one of the last 14 commits is a direct push to `main`** (CI event `push`,
-branch `main`). CI is green on all of them — so this is currently latent, not active, harm.
+`AUDIT_2026-09-01` §D5 called branch protection *"the most important item in this document"*.
+**It is applied** (checked through the API, 2026-09-14): `main` requires all six CI checks, strictly
+(the branch must be up to date), and refuses force pushes. What it does **not** do is apply to
+admins (`enforce_admins: false`) or require a review — which is why every one of the last 14
+commits is a direct push to `main` (CI event `push`) and still lands. CI is green on all of them, so
+this is latent, not active, harm.
 
-It stops being latent the moment a second contributor arrives, which is precisely what the Reddit
-announcement is for.
+For anyone else the gate is already real: an outside pull request cannot merge red. The open
+question is only whether the owner's own pushes should go through it too — one API call
+(`POST /repos/superwilso/Cinder/branches/main/protection/enforce_admins`), and after it every change,
+the owner's included, needs a pull request with green checks.
 
 ### A5. The highest-consequence unverified code in the tree is shipping to strangers right now
 
@@ -94,8 +98,14 @@ executed on Windows"* — and since Sony's updater was removed, **it is the only
 can install anything.** It is live in v0.3.4, which is what a Reddit link points at.
 
 Downloads as of this audit: **2 Windows, 1 Linux.** The window to test it before it meets an
-audience is open and small. `DEVICE_CHECKLIST.md` 11.9 already describes the test; it is re-opened
-and unrun.
+audience is open and small. `DEVICE_CHECKLIST.md` 11.9 already describes the test.
+
+> **Update, later on 2026-09-14 — owner-reported, no log captured:** the owner has run the
+> **v0.3.3** installer. `installer/` is byte-identical between `v0.3.3` and `v0.3.4` apart from the
+> version number (`git diff --stat v0.3.3 v0.3.4 -- installer/`: `Cargo.toml` and `Cargo.lock`, one
+> line each), and the Windows pass-through landed before `v0.3.2`, so a v0.3.3 run on Windows
+> exercises exactly the code v0.3.4 ships. If that run was the Windows executable, A5 is closed in
+> substance; 11.9 still wants its install log attached to be recorded as a pass.
 
 ---
 
@@ -107,10 +117,10 @@ Ordered by *consequence × cheapness*, not by interest.
 
 | # | Action | Why now | Cost |
 |---|---|---|---|
-| 1 | **Run `DEVICE_CHECKLIST.md` 11.9** — the Windows installer, end to end, from the GitHub release | The only install path for Windows users has never run once (A5). A stranger's first experience is currently untested code | One device session |
+| 1 | **Record `DEVICE_CHECKLIST.md` 11.9** — the Windows installer, end to end, from the GitHub release | The only install path for Windows users. The owner has since run the v0.3.3 installer (A5, update); what is missing is the platform and the install log that make it a recorded pass | One device session |
 | 2 | **Banner `CLAUDE.md`**, pointing at `STATUS.md` / `DEVICE_CHECKLIST.md` / this file | It is read first by every session and is seven weeks wrong (A2) | Minutes |
-| 3 | **Turn on branch protection** on `main` (§D5's API call, contexts updated for today's job names) | A public repo with no merge gate, about to receive its first outside PRs (A4) | One API call, needs the owner |
-| 4 | **Point the Reddit post at GitHub Issues** | Templates already exist (`bug_report`, `device_report`, `feature_request`) and route correctly. It is also the only channel this session can watch | One post edit |
+| 3 | **Decide whether branch protection binds the owner** (`enforce_admins`) | Protection is on and already gates outside PRs; the owner's direct pushes bypass it (A4). Owner's call — it turns every change into a pull request | One API call, needs the owner |
+| 4 | **Point the Reddit post at GitHub Issues** | Templates already exist (`bug_report`, `device_report`, `feature_request`) and route correctly. It is also the only channel a cloud session can watch (D1) | One post edit |
 
 ### B1. Next — cheap, and each one removes a whole class of future pain
 
@@ -288,20 +298,24 @@ and the palette loader both follow). This extends that from "unknown" to "unsurv
 ## Part D — Maintainer watch
 
 **Decisions, 2026-09-14:** fixes go to a **draft PR** and never straight to `main`; **any** new
-issue or bug report pings the owner; **GitHub is the channel** (Reddit is unreachable — see below);
-device work happens by **script-and-paste**.
+issue or bug report pings the owner; **GitHub is the channel** (a cloud session cannot read Reddit —
+see below); device work happens by **script-and-paste**.
+
+### D1. Reddit cannot be read from the cloud build environment — it can from the owner's machine
 
-### D1. Reddit cannot be read from the build environment
+In the cloud session that wrote this, tested four ways: `WebFetch`, `curl https://www.reddit.com/...`,
+`old.reddit.com`, and the `.json` API. All four fail identically at the proxy with
+**`connect_rejected` — 403 to CONNECT (organization policy)**: an egress allowlist denying the
+domain, which RSS shares. No workaround was attempted there, deliberately.
 
-Tested four ways: `WebFetch`, `curl https://www.reddit.com/...`, `old.reddit.com`, and the `.json`
-API. All four fail identically at the proxy with **`connect_rejected` — 403 to CONNECT
-(organization policy)**. This is an egress allowlist denying the domain, not rate limiting and not
-a User-Agent problem; RSS resolves to the same blocked host and fails the same way. `WebSearch`
-does not surface the thread either.
+**That is a property of that environment, not of Reddit.** On the owner's own machine the thread's
+Atom feed (`…/comments/1wedu23/.rss?sort=new&limit=100`, browser User-Agent) returns 200, and a
+local session has been watching it alongside GitHub on 2026-09-13/14. The `.json` endpoints do 403
+there too, and WebFetch is blocked for the domain, so the feed is the one route.
 
-No workaround was attempted, deliberately: routing around an explicit network policy is not a thing
-to do quietly. **The supported path is to funnel reports to GitHub Issues**, where the templates
-already exist and already route bricks/boot failures to `RECOVERY.md` and `SECURITY.md`.
+So the conclusion stands for anything **scheduled in the cloud**: funnel reports to GitHub Issues,
+where the templates already exist and already route bricks/boot failures to `RECOVERY.md` and
+`SECURITY.md`. A local session can keep reading the thread directly.
 
 ### D2. What the watch does
 
@@ -312,11 +326,14 @@ An hourly Routine wakes a session that:
 2. for anything new: reproduces it if it can be reproduced off-device, writes the fix, runs
    **all** gates (549 tests, 41 harness scenarios, syntax check, `golden.txt`), pushes a branch and
    opens a **draft PR**;
-3. replies on the issue with the diagnosis and the PR link;
-4. **pings the owner's phone** — on every new report, per D's decision, whether or not a fix landed;
+3. **drafts** a reply with the diagnosis and the PR link, and hands it to the owner — it does not
+   post it. Anything said in public on the project is in the owner's own words;
+4. **pings the owner's phone** — on every new report, per D's decision, whether or not a fix landed,
+   with the draft reply attached;
 5. if it is device-gated, writes a self-contained diagnostic script, says so, and does not guess.
 
-**The boundary it does not cross:** nothing merges to `main`; nothing is pushed that fails a gate;
+**The boundary it does not cross:** nothing merges to `main`; nothing is posted publicly in the
+owner's name; nothing is pushed that fails a gate;
 anything touching the boot path, the installer, root helpers or Sony IPC gets a diagnosis and a
 proposed patch rather than a speculative fix, because CI cannot execute that half of the tree
 (A1) and a wrong guess there costs a device with no public recovery path.