v0.14.8

fix(chatgpt): tolerate null catalog windows and fetch once per account

The plan catalog may send `context_window: null`; parsing it as `usize`
failed the whole catalog, so model listing, account selection and the
startup window resolution all broke for that account. Read it as optional
and treat a missing value as unknown.

Startup and model reload resolved legacy zero windows per provider, and
every earlier model selection leaves its own enabled provider, so one
account fetched the same catalog serially once per model. Resolve windows
in `completion_models`, fetching each account's catalog at most once;
`completion_model` is synchronous again.

Update anda_kip to 0.14.5 (documentation and schemas only) and
anda_cognitive_nexus to 0.14.5, which replaces the cognitive-memory and
general-domain revisions bundled by 0.14.0-0.14.4 when a Nexus opens. Both
resolve to a single version across Bot, Brain and Engine.

Validation: full anda_bot suite (1526 passed, 2 ignored), Clippy
all-targets/all-features, formatting and diff checks clean.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>