Render integration capability descriptions in the execute inventory

[justcarlson]

Problem

The execute tool description and the execute skill list connected integrations as bare slugs (packages/core/execution/src/description.ts said "no descriptions" by design). An agent opening a fresh session cannot tell what each integration is for. On a deployment with 40+ integrations, models burn sandbox calls on discovery, or guess native paths that do not exist. Reported from production self-host use in justcarlson/Silent-Cartographer#184 and related to the catalog-ergonomics requests in #2130.

Change

Each inventory line now renders the integration's catalog description:

- `seerr` — request movies/TV into Radarr/Sonarr
- `github`

• One scannable line per integration: first line of the description, whitespace collapsed, capped at 120 chars on a word edge. The block is always-loaded prompt context, so the line budget stays tight.
• Legacy rows whose description merely repeats the slug or display name (case-insensitive) stay bare, so existing catalogs do not get noisier until curated.
• Descriptions stay user-editable data via PATCH /api/integrations/:slug; deployments tune wording without a rebuild.
• parseIntegrationInventory tolerates the optional — description suffix, so hosts deriving search_<integration> tools from the built description parse the same slugs.

One extra executor.integrations.list() per session build, beside the existing connections.list().

Tests

• Updated: the "names only" test now pins description rendering and slug/name-repeat suppression.
• Added: truncation to one line (long, multiline descriptions).
• Added: parseIntegrationInventory reads slugs past description suffixes.
• packages/core/execution: 80/80 pass. packages/hosts/mcp: 251/251 pass. oxlint and tsgo clean.