diff --git a/docs/README.md b/docs/README.md
index 29d550317..d41a85c3d 100644
--- a/docs/README.md
+++ b/docs/README.md
@@ -74,6 +74,7 @@ A document in `docs/` proper should be one somebody keeps current.
 
 ### Integrity and integrations
 
+- [`dependency-advisories.md`](dependency-advisories.md) — open Dependabot alerts whose patched version is unreachable from this dependency graph: the resolver refusal, the exposure assessment, and the upstream change that would lift each
 - [`project-permission-integrity.md`](project-permission-integrity.md) — reconciling stale grants against the project registry
 - [`discord-rich-presence.md`](discord-rich-presence.md) — privacy-safe local activity publishing
 
diff --git a/docs/dependency-advisories.md b/docs/dependency-advisories.md
new file mode 100644
index 000000000..42c20d328
--- /dev/null
+++ b/docs/dependency-advisories.md
@@ -0,0 +1,174 @@
+# Dependency advisories with no available fix
+
+Most Dependabot alerts close with a version bump. A few cannot, because the
+patched version is unreachable from this dependency graph — a transitive crate
+is pinned by an upstream generation that has not moved, or has stopped moving
+altogether. Those alerts stay open indefinitely, and every session that finds
+one re-derives the same constraint from scratch before concluding there is
+nothing to do.
+
+This file is where that derivation gets written down once. An entry records
+four things and nothing else:
+
+1. **The constraint** — the exact resolver refusal, quoted, not paraphrased.
+2. **The evidence** — the commands that produce it, so the next reader can
+   re-run them rather than trust this page.
+3. **The exposure** — whether the vulnerable code path is reachable from what
+   this repo actually builds and ships.
+4. **The condition that lifts it** — the specific upstream change that would
+   make a bump possible, so the entry can be retired deliberately.
+
+An entry is deleted when its alert closes, not edited into a historical note.
+The alert is the source of truth for state; this page only explains it.
+
+---
+
+## Dependabot #18 — `glib` `VariantStrIter` unsoundness
+
+| Field | Value |
+|---|---|
+| Alert | [#18](https://github.com/OpenCoven/coven-cave/security/dependabot/18) |
+| Advisory | [GHSA-wrw7-89jp-8q8g](https://github.com/advisories/GHSA-wrw7-89jp-8q8g) (no CVE) |
+| Severity | Moderate |
+| Package | `glib` (crates.io), transitive |
+| Manifest | `src-tauri/Cargo.lock` |
+| Vulnerable | `>= 0.15.0, < 0.20.0` |
+| Patched | `0.20.0` |
+| Resolved here | `0.18.5` |
+| Status | **Not fixable at this Tauri generation.** No exposure identified. |
+| Assessed | 2026-08-19 |
+
+### What the advisory is
+
+`VariantStrIter::impl_get` passed `&p` — an immutable reference to a
+`*mut libc::c_char` initialized to `NULL` — as a variadic out-argument to
+`glib_sys::g_variant_get_child`, which mutates the pointer in place. The
+mutability mismatch went uncaught because the C function is variadic. Recent
+Rust compilers disregard the unsound write when building with optimizations, so
+`CStr::from_ptr` then receives `NULL` and the process dereferences it.
+
+The failure mode is a **crash**, not a disclosure or code-execution primitive,
+and it only occurs on a code path that iterates a string-array `GVariant`.
+
+### The blocking constraint
+
+`glib` is not in `src-tauri/Cargo.toml`. It arrives entirely through the
+Linux/BSD GTK3 stack that `tauri` and `wry` depend on:
+
+```console
+$ cd src-tauri
+$ cargo tree -i glib --target x86_64-unknown-linux-gnu
+glib v0.18.5
+├── atk v0.18.2
+│   └── gtk v0.18.2
+│       ├── libappindicator v0.9.0
+│       │   └── tray-icon v0.23.1
+│       │       └── tauri v2.11.2
+│       ├── muda v0.19.2
+│       ├── tao v0.35.3
+│       │   └── tauri-runtime-wry v2.11.2
+│       ├── tauri v2.11.2
+│       ├── tauri-runtime v2.11.2
+│       ├── webkit2gtk v2.0.2
+│       │   └── wry v0.55.1
+│       └── wry v0.55.1
+├── cairo-rs v0.18.5
+├── gdk v0.18.2
+├── gdk-pixbuf v0.18.5
+├── gdkx11 v0.18.2
+├── gio v0.18.4
+├── gtk v0.18.2
+├── javascriptcore-rs v1.1.2
+├── libappindicator v0.9.0
+├── pango v0.18.3
+├── soup3 v0.5.0
+└── webkit2gtk v2.0.2
+```
+
+A plain `cargo update -p glib` reports `Locking 0 packages` — `0.18.5` is
+already the newest release in the `^0.18` range. Forcing the patched version
+names the pin directly:
+
+```console
+$ cargo update -p glib --precise 0.20.0
+error: failed to select a version for the requirement `glib = "^0.18"`
+candidate versions found which didn't match: 0.20.0
+location searched: crates.io index
+required by package `gtk v0.18.2`
+    ... which satisfies dependency `gtk = "^0.18"` (locked to 0.18.2) of package `tauri v2.11.2`
+    ... which satisfies dependency `tauri = "^2.11.2"` (locked to 2.11.2) of package `app v0.3.7`
+```
+
+### Why upgrading Tauri does not help
+
+The pin is not this project's Tauri version. It is the GTK3 Rust binding
+generation, which has stopped:
+
+- **`gtk` tops out at `0.18.2`**, published 2024-12-09. That is the highest
+  version the crate has ever had. `glib 0.20` belongs to the gtk4-rs
+  generation; there is no GTK3 binding release that consumes it.
+- **`webkit2gtk 2.0.2`** — the newest, and pinned exactly as `=2.0.2` by `wry`
+  — requires `glib ^0.18.0`, `gtk ^0.18.0`, `gio ^0.18.0`.
+- **`tauri 2.11.5`**, the latest release at time of writing, still declares
+  `gtk ^0.18` for `cfg(any(target_os = "linux", …bsd))`. So does
+  `wry 0.56.1`.
+
+Every Tauri v2 release therefore resolves `glib` to the `0.18` line on Linux.
+Bumping this project's `tauri` from `2.11.2` to `2.11.5` would change nothing
+about this alert.
+
+### Exposure
+
+**Linux/BSD only.** The whole GTK stack sits behind
+`cfg(any(target_os = "linux", target_os = "dragonfly", target_os = "freebsd",
+target_os = "openbsd", target_os = "netbsd"))`. It is absent from the macOS and
+Windows builds, which is what `tauri.conf.json` bundles (`dmg`, `app`, `msi`).
+The Linux artifact is real, though — `.github/workflows/release.yml` builds an
+AppImage on `ubuntu-22.04` — so "Linux-only" narrows the blast radius rather
+than eliminating it.
+
+**No reachable call site was found.** `VariantStrIter` is only constructed by
+`glib::Variant::iter_str()`. Scanning the vendored sources of every crate in
+the tree that depends on `glib`:
+
+```console
+$ grep -rl 'iter_str\|VariantStrIter' ~/.cargo/registry/src/*/<crate>/
+```
+
+returns matches in `glib-0.18.5` itself and in **none** of `gtk-0.18.2`,
+`gdk-0.18.2`, `gdkx11-0.18.2`, `gdk-pixbuf-0.18.5`, `gio-0.18.4`, `atk-0.18.2`,
+`cairo-rs-0.18.5`, `pango-0.18.3`, `soup3-0.5.0`, `javascriptcore-rs-1.1.2`,
+`webkit2gtk-2.0.2`, `libappindicator-0.9.0`, `muda-0.19.2`, `tray-icon-0.23.1`,
+`tao-0.35.3`, `wry-0.55.1`, `tauri-2.11.2`, `tauri-runtime-2.11.2`, or
+`tauri-runtime-wry-2.11.2`.
+
+`src-tauri/src` contains no `glib`, `gio`, or `Variant` usage of its own.
+
+Assessed exposure: **none identified**, against a worst case of a null-pointer
+crash in the Linux AppImage.
+
+### What would lift this
+
+Either of:
+
+- **`wry` moving its Linux webview off GTK3.** A `webkit2gtk-6.0` / GTK4
+  generation would bring `glib 0.20+` with it. `wry` pins `webkit2gtk =2.0.2`
+  exactly, so this is visible as a change to that pin — watch it rather than
+  watching `glib`.
+- **An upstream backport to the `glib` 0.18 line.** The 0.18 series is not
+  receiving fixes, so treat this as unlikely.
+
+Nothing in this repository can accelerate either one. Do **not** hand-edit
+`Cargo.lock` to a `glib` version the resolver cannot reach; the build would
+fail and the alert would not close.
+
+### Re-checking
+
+```bash
+cd src-tauri
+cargo tree -i glib --target x86_64-unknown-linux-gnu
+cargo update -p glib --dry-run --precise 0.20.0   # expect the refusal quoted above
+```
+
+If that refusal no longer names `gtk v0.18.2`, the constraint has moved and
+this entry should be re-derived rather than trusted.