feat(nodes): 876 show a pgpu card's device profile in the row and the full view

[raven-pan]

What type of PR is this?

/kind feature

What this PR does / why we need it

A pgpu (passthrough) card is scheduled by name: a flavor asks for it with
accel:device_profile=<name>, and Cyborg matches that name to the card. Until
now the UI showed that name nowhere, so an operator had to read it off the
controller before they could write the flavor.

The card now carries the name in both places it can appear:

• The expanded row shows it under Profiles / ID, with a copy button.
• The full-view modal's Profiles / ID tab shows it too. The tab used to be
  locked for a pgpu card.

Two bugs found on the way came along:

• The Profiles / ID and Attached Instances tables sat 32px from their
  border, twice what the design asks for.
• A failed Edit GPU Type wrote the reason to the console only, so the modal
  looked the same before and after pressing Confirm.

Screenshot placeholders — attach by hand

Full-view modal, pgpu card, Profiles / ID tab

Expanded row, both tables with the corrected padding

Full-view modal, MIG-backed vGPU, unchanged

Edit GPU Type modal, confirm step, API error shown

---

QA perspective

What changed on screen

Node Details page > GPU Resources table.

Place	Before	After
Expanded row of a Passthrough card	no Profiles / ID section	Profiles / ID (1) with an Alias Name column, the device profile name and a copy button
Full-view modal of a Passthrough card	Profiles / ID tab locked, opens on Attached Instances	tab opens first and holds the same name, one Alias Name column, one row, no copy button
Padding of Profiles / ID and Attached Instances	32px from the border to the first cell	16px, the same as the gap between two cells
Edit GPU Type when the API rejects the change	nothing on screen, modal stays open	red message with the API text above the step content, modal stays open

A Passthrough card whose device profile is missing keeps the old behaviour: no
section in the row, Profiles / ID tab locked, modal opens on
Attached Instances.

SR-IOV vGPU and MIG-backed vGPU cards keep the tables they had. Only their
padding changes.

Repo dependency

• cube-cos-api supplies the deviceProfile field on
  GET /nodes/{node}/gpuCards. Without an API build that carries it, the field
  arrives as null and the section stays hidden — that is the expected
  fallback, not a defect.
• cube-cos-openapi holds the contract. The field is already on develop, so
  this PR moves no submodule.
• cube-cos-ui (this repo) renders it.

Check list

1. Open Node Details for a node with a Passthrough GPU attached to an instance.
2. Expand the Passthrough card's row. Profiles / ID (1) shows, column header
   Alias Name, the device profile name below it.
3. Click the copy button. The clipboard holds the exact name.
4. Paste it into
   openstack flavor set --property accel:device_profile=<name>. The flavor
   schedules onto that card.
5. Click the full-view button on the same row. The modal opens on
   Profiles / ID and shows the same name, one column, one row, no copy button.
6. Switch to the Attached Instances tab and back. Both tabs stay clickable.
7. Expand an SR-IOV vGPU card. Its profiles table keeps the VRAM, Counts,
   Alias Name columns.
8. Expand a MIG-backed vGPU card. Its profiles table keeps the extra
   Remaining column. Open its full view — five columns and pagination, as
   before.
9. Measure any expanded table: the border to the first cell is 16px, the same
   as the gap between two columns.
10. A Passthrough card in Idle state stays non-expandable.
11. Open Edit GPU Type on a card the API will reject — for example a card an
    instance still holds. Press Next, then Confirm. A red message shows the
    API's own text above the confirm table. The modal stays open and Confirm
    stays clickable.
12. Press Back to edit in that state. The message disappears. Close and
    reopen the modal — no stale message.
13. Switch the UI language to 中文. Titles and column headers are translated;
    the profile name and the API error text are not.

Developer perspective

The device profile (fb6879e6)

• GpuDeviceProfile.tsx — new component. It returns null on a falsy
  deviceProfile, so it renders on a pgpu card only. The API reports null for
  every other resource type, and for a pgpu card whose profile does not exist or
  could not be read; it gives no reason for the absence, so a label with no
  value would tell the operator neither what to paste nor why.
• The block renders through GetInfoTable, the same component GpuDetails.tsx
  uses for the profiles and instances tables, with scrollBehavior="vertical"
  and one row. Title and column header reuse
  nodes.details.profilesIdList.title and
  nodes.details.profilesIdList.aliasName.
• The section reuses existing keys, so the JSON resources gain none. An earlier
  draft added nodes.details.deviceProfile.title; it is not in this branch.
  If that row reached the translation sheet, delete it there, or the next
  pnpm i18n:sync brings it back.
• FullViewDeviceProfileTable.tsx — new component. One Alias Name column
  through GetCosBasicTable, no pagination: a pgpu card has exactly one device
  profile.
• GpuDetailsFullViewModal.tsx — the tab is disabled only when a pgpu card has
  no deviceProfile, not for every pgpu card. The default tab follows the same
  flag, so the modal opens on Profiles / ID whenever that tab has content.

The padding (f9041507)

InfoTable.tsx set border-separate border-spacing-4 on the table and p-4 on
the box around it. border-spacing also applies at the table's outer edge, so
the two stacked: 16 + 16. Dropping the box padding leaves one source of
spacing.

Measured against the design (Figma nodes 130:21557 and 130:21598): border to
first cell 32 -> 16. Column gap, header to first row and title to table were
already 16 / 16 / 8.

The change reaches four call sites, all under NodeGpuResources: the two tables
in an expanded row, the pgpu block, and the confirm step of the Edit GPU
Resource modal.

The error message (5c4ea995)

EditGPUResourceModal.tsx only wrote the failure to the console, and its catch
does not close the modal, so the confirm step stayed on screen with no reason.
It now renders errorState in a CosInlineNotification above the step content.

Inline, not a toast: the surface the action came from is still on screen, and
the operator can press Confirm again. That matches VerifyEmailSenderModal and
the tunings / triggers forms. A toast fits an action whose surface disappears,
such as GpuConsoleLink.

useCosMutationRequest already returns errorState and clearError, so the
modal holds no state of its own. oversee clears the error when the next
request starts; clearError covers Back to edit and closing the modal.

Mocks

src/mocks/gpu.ts already carries deviceProfile on every card from develop.
handlers.ts is unchanged on this branch — register mockListNodeGpuCards
locally to see the cards without a GPU host.

Which issue(s) this PR fixes

Fixes bigstack-oss/cubecos#876

Special notes for your reviewer

• The PR carries three commits: the device profile, the padding, the error
  message. The last two came out of reviewing the first.
• The branch was rewritten twice: a commit that carried only local scratch files
  (.claude/settings.json, .playwright-mcp/, two throwaway .mjs probes) is
  gone -- its message claimed an openapi submodule bump it never contained --
  and the rest squashed into the two commits above.
• pnpm test: 7 files, 121 tests pass.
• pnpm lint: tsc clean for every package. Prettier flags
  .impeccable/hook.cache.json, a local tool cache outside this branch.
• One 1px gap is left: the design puts 32px between rows, the UI renders 31,
  because primary-body5 has a 15px line height. Not addressed here.
• Screenshots must be attached by hand.

Additional documentation

🤖 Generated with Claude Code