test(enterprise): cover the model availability round trip (#1669)

[rafaelgiln]

Closes #1669. The write half of the models screen, which #1660 excluded on a premise that turned
out to be my own measurement error.

1. Edit models works — #1659 was three errors, not a defect

It is a toggle: aria-pressed goes false → true, and each of the 44 rows swaps its state
text for a control.

view mode:   - text: Disabled
edit mode:   - switch "Show gpt-5.6-sol in model pickers"

with Enable all models, Hide all models, Cancel and Save models appearing alongside.

	Models list	items	buttons	panel text
before	1	44	6	1645 chars
after	1	44	9	1295 chars

The three errors, recorded in the doc because the same three would catch anyone probing this
screen:

1. Looking for a dialog — role=dialog, role=menu, a URL change, all correctly zero, because
   none of them is how the control works.
2. Reading the shrinking panel as a collapse — 1645 → 1295 is not the list disappearing; the
   item count is 44 in both states. It is 44 rows each losing the word Disabled to a switch.
3. Counting the wrong role, twice — role=switch before entering edit mode (0, correctly, and
   the probe's own guard then meant it never clicked at all), then role=button with the toggle's
   name where the control is a role=switch.

That is the third time in this area a partial read of a screen produced a confident wrong
conclusion — after the delete button's aria-label and the provider card that "vanished" while
sitting in the next region of the same screen. #1659 is closed invalid with the same write-up.

2. Covered tests

#	Test	What it validates
1	entering edit mode offers one control per model and writes nothing yet	One switch per model, each naming its model; enabled_model_keys untouched. A screen that wrote on entering an edit mode would be surprising in the worst way.
2	enabling a model and saving lands exactly that model in the policy	Flipping stages with no request; Save models issues PUT /api/v1/model-availability-policy → 200; the policy holds the key of the model the switch named, not a position.
3	enabling a model does not make it visible while the provider has no credentials	The row stays Pending / Hidden.
4	cancelling an edit discards it	No write leaves the browser, asserted on the request.

3. Test 3 replaced the test I was about to write

The natural assertion — "the row stops reading Hidden" — is false. Measured: after saving one
enabled model the row still reads Pending / Hidden, and that is correct. Builder visibility
follows the provider's credentials, not the availability policy: one enabled model on a provider
that is still Pending is not something a builder can pick.

Asserting the natural version would have manufactured a defect. The two are different axes and the
screen does not conflate them — the same lie #1660's read tests catch, arriving from the other
direction.

4. Force-fail

Mutation	Result
edit mode never entered	red (4 of them)
the saved key asserted as a model the test did not pick	red
the row asserted Visible after enabling	red
Cancel replaced by Save models	red

5. Validation

langflow-enterprise:latest (2026-08-27), RBAC variant.

• Spec alone: 8 passed (4 read + 4 write), ~29 s.
• Whole admin-console/: 29 passed, twice, ~1.3 min.
• Instance left clean: enabled_model_keys: [], no governance records.

One advisory, noted and deliberately not filed. 503 /api/v1/session — "Authorization service
unavailable" — appeared once across the two directory runs. Not reproducible: 76 polls of that
endpoint during four consecutive policy writes all answered 200, and the same shape has been
observed once before on this image without reproducing. The fixture logs it as advisory and does not
fail, which is the right handling for a single unreproducible transient — filing it would send
someone after a phantom, and hiding it would be worse.

6. Notes for the reviewer

• Teardown unchanged from enterprise admin: no test covers the provider-approval screen, where one click decides what the installation may use #1656: DELETE /api/v1/model-provider-governance/{id} with
  expected_revision, which removes the approval and the availability it created. Each test revokes
  before it approves rather than assuming a clean instance.
• No @stable (project-management: the destructive 'deleting every folder' test wipes sibling tests' folders and flows under fullyParallel #1010). Tags unchanged.
• QA-CHECKLIST.md § 22.7: the [!] bullet pointing at enterprise models screen: Edit Models is enabled, opens nothing, and collapses the row instead #1659 becomes four [-].

🤖 Generated with Claude Code