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.
Keep only groups with another path in the active round. This bounds skewed sampling by paths visited instead of rescanning every singleton for every late-group item.
boot() asserts over ctx.loader after awaiting the Loader, but the TUI renders
as soon as its own fiber starts: an /exit typed before the last entry settles
runs disposeRootAndExit, which takes the Loader service with the tree. The
assertions then read undefined and crashed the process with a TypeError over
an app that exited exactly as asked. The keyless personal-overlay PTY smoke
lost this race in roughly two of three runs.
The web_search execute result and searchMetaFromValue each spread the
same {url, title?, snippet?, publishedAt?} projection over a seam source,
which the duplication gate flags as a clone. Extract projectSource, typed
on the seam's WebSearchSource, so both sites carry a byte-identical shape
from one definition.
The rename to deepseek-official reached cordis.snapshot.yml but not its
persistent-tools sibling, so the replay catalog no longer owned the route the
agent asks for. The SDK server then mounted the real adapter, which failed the
turn on a missing key. Re-records the six transcripts and the two diagnostics
cards that still carried the old route name.
renderFetchOutput has no external consumer: only formatFetchOutput and
fetchMetaFromValue call it, both through the registry, which deep-freezes
the result value. Exporting it let a hypothetical caller mutate a cached
input or the returned RenderedFetch and desync the card's truncated flag
from the model text. Drop it from the barrel and document that the memo
needs no defensive copy because every caller is internal and read-only.
The mutate seam's root-path ops, its non-array rejection, and the recursion
into an existing nested object were unexercised, as was the whole containment
side of the settings/document-updated fan-out and the before-snapshot repair
of a section a hand edit left non-object. The settings.mutate route had no
round trip at all: neither the client method nor the handler entry ran.
