fix: make overview tokens readable down the column

[yujiezhang-ops]

Four things the Agent overview got wrong, all in how a row presents state.

Token order was not stable

The Provider token was dropped when empty, which slid model and version left — so the same field sat in a different slot from row to row and the strip could not be read down the column.

Order is now fixed, Profile → Provider → model → version, with the first two slots always rendered. A new test asserts the order with a field absent, which is the case that used to shift.

"Not linked" / "Not configured" were invented vocabulary

Worse, 未绑定 / Not linked was shown for both a missing Provider and a missing Profile, so two identical pills gave the user no way to tell which piece was actually absent.

Each token now names its own field — 无 Profile / 无 Provider — following the existing Provider 缺少 Key → Provider key missing wording rather than introducing a term. The separate 未配置 badge is gone; it said the same thing a third time.

targetSummary() still returns 未配置 internally, but line 240 suppresses it, so it never reaches the user. Left as-is since it answers a different question ("the detected config is empty"), though using a translated string as a sentinel is fragile and worth replacing later.

Configuring an installed Agent was hidden

配置 appeared in the action row only when the Agent could not launch; for an installed one it was reachable only by opening <details>. That made the common case the hidden one. It now sits with 更新 and 启动, and the duplicate link inside <details> is removed.

The desktop card did not match the rows beneath it

Four labelled fact columns at min-height: 145px — twice the height of the CLI rows for the same four values. Now the same token strip, same order, same wording, same 84px.

Trademark marks, restored with an auditable basis

Claude Code, Cursor and Kilo CLI are real artwork again, taken from the MIT-licensed lobehub/lobe-icons already vendored in this repo rather than by restoring the unaudited vendor favicons that #25 removed.

Each is registered in asset-rights.json with source URL, licence, copyright owner and SHA-256; the licence bundle is regenerated and --check passes. Aider has no mark upstream (HTTP 404), so it keeps a generic symbol rather than gaining a favicon with no redistribution basis.

agents.test.tsx previously asserted rights for one example (["codex"]). It now asserts the whole asset set, so an unregistered mark fails there. Verified by mutation: flipping one license to UNKNOWN fails the test.

Incidental cleanups

A leftover duplicate .agent-manage-identity rule set display: grid with no columns and won on source order, stacking each Agent's name under its icon — that was the actual cause of the tall, narrow rows. Also removes CSS orphaned by the changes above (.desktop-app-fact, .agent-manage-detail-action) and ignores __pycache__, which the Python unittest run leaves behind untracked.

Two dead selectors (.desktop-app-path, .agent-manage-target) predate this branch and are left alone.

Verification

Measured in the running app, not inferred from the diff:

token order   配置模板 → 模型服务 → 模型 → 版本   (CLI and desktop identical)
actions       配置 / 更新 / 启动
row height    112px → 89px
old wording   未绑定 / 未配置 absent from the page

pnpm run test                  148 passed (22 files)
pnpm run build                 typecheck + vite ok
go vet ./...                   clean
generate_third_party_licenses.py --check   ok
check-docs.py                  ok: 49 files

🤖 Generated with Claude Code