fix(llm,settings): refuse post-disposal route replacement and teardown notifications

Two lifecycle holes the registry and the consumer helper left open.

`AdapterRegistrationHandle.replace` had no liveness guard: after the handle's
disposer ran, a replace put routes back into the registry with nothing left to
release them, so the adapter leaked permanently. `owned` being empty cannot
carry that fact, because `replace([])` is the legal empty-section state, so the
disposer records it explicitly.

`installSettingsSection`'s watcher lacked the guard its own disposer carries:
a stored change landing while the consumer unloads reached `onChange`, which
re-registers routes against a fiber whose resources are being released.

Also documents `withFileLock` in the atomic-write README (it claimed one
export), records the age-based lock takeover as a known limitation, and lists
ctx.settings and ctx.credentials in the architecture capability table.

packages/util/atomic-write/README.md

@@ -7,14 +7,20 @@ Zero-dependency atomic file replacement shared by file-backed stores that must n
 ## Surface
 
 ```ts
-import { writeFileAtomic } from '@deepseek-ai/dsh-atomic-write'
+import { withFileLock, writeFileAtomic } from '@deepseek-ai/dsh-atomic-write'
 
 declare const text: string
+declare const render: (previous: string) => string
 
 await writeFileAtomic('/home/u/.dsh/settings.yaml', text, { mode: 0o600 })
+
+// Read-modify-write against the same file from several processes.
+await withFileLock('/home/u/.dsh/settings.yaml', async () => {
+  await writeFileAtomic('/home/u/.dsh/settings.yaml', render(text), { mode: 0o600 })
+})
 ```
 
-One export. The contract, in the order failures would exploit it:
+`writeFileAtomic` commits one already-rendered string. The contract, in the order failures would exploit it:
 
 - **Exclusive-create temp** (`wx`, random suffix): the open refuses to follow a symlink planted at a guessable temp path.
 - **The fresh inode carries `mode` through the rename**: replacing a wider-permission file narrows it without a chmod race. `mode` is required so the permission decision stays visible at every call site (subject to the process umask, like every fresh inode).
@@ -22,6 +28,8 @@ One export. The contract, in the order failures would exploit it:
 - **Same-directory sibling** keeps the rename on one filesystem, so the swap stays atomic.
 - Parent directories are created; on any failure the temp is removed and the failure rethrown; readers observe either the old or the new complete content.
 
+`withFileLock` serializes the writers of one file across processes, for the read-render-commit cycles a bare atomic commit cannot make safe on its own. The lock is a `wx`-created `<filename>.lock` sibling, so readers never contend; waiters back off exponentially and fail with a timeout rather than block forever. A lock older than the stale age is treated as a crashed holder and broken — see [Known Limitations and Deferred Work](#known-limitations-and-deferred-work) for what that costs.
+
 ## Model Experience
 
 None, as this is a pure filesystem primitive; nothing here reaches a model request.
@@ -34,3 +42,4 @@ None; nothing here enters a request prefix.
 
 - **Atomic, not durable** — no `fsync` of the file or its directory, so after a crash the rename may be observed unwound. The file-backed stores here re-read and republish on boot, keeping durability the caller's policy.
 - **String content only** — no `Buffer` or stream form until a consumer needs one.
+- **The lock takes over by age, not by ownership** (`TODO(settings-lock-ownership)`) — a holder slower than the stale age has its lock broken by a waiter, and release unlinks the path unconditionally, so a slow writer can remove a successor's lock. Two writers can then overlap and one cycle's result be lost. The stale age is set well above any write this repo performs, so the exposure is a paused or swapped-out process; ownership-safe acquisition and release is the fix.