Main is red on four counts right now, none of them caught before landing because each is a test that reads source text or a shape rather than behaviour. Every open PR that runs these jobs inherits all four, and merging main into a branch cannot clear them.

1. The mmproj fallback import has no file extension

chat/utils/image-input-support.ts gained import { isTextOnlyMmprojFallback } from "./mmproj-fallback". Before #9173 that module held only import type lines, which are erased, so it had no runtime relative import at all. node --experimental-strip-types cannot resolve an extensionless specifier, so queued-model-capabilities.test.ts fails on a pristine checkout. Adds the .ts.

This is the one that makes Frontend build + bundle sanity and Frontend unit tests (Windows) fail on every PR.

2. The autoload harness has no stub for mmprojFallbackMessage

test_chat_autoload_failure_gate.py slices autoLoadSmallestModel out of chat-adapter.ts and runs it. Its own guard fails the whole file when the slice references a name the preamble does not define, and #9173 added such a name. 71 failures.

3. The picker contract pinned one spelling of the success toast

test_model_picker_contracts.py asserted the literal description: cpuFallbackReason. The mmproj branch now sits ahead of the CPU one in that ternary. Pinned on the description expression instead of which fallback is tested first, so the next branch to land does not break it again.

2 and 3 are what make Repo tests (CPU) fail.

4. _strip_mmproj_args lost its old argument name

test_tensor_split_abort_raises_early_to_layer_fallback pins the order of two branches in load_model by their source offsets. #9173 renamed the variable handed to _strip_mmproj_args, so find() of the old spelling returned -1 and the ordering assert quietly compared against it. Pins the call itself and fails loudly when it is no longer reachable. This is (Python 3.13).

5. The chat UI smoke races the lazy route's Suspense fallback

Separate root cause, same symptom of an unreliable red. exercise_permission_mode_controls asserts the permission pill on expect's 5s default. Four of its six assertions come straight after page.reload(wait_until = "domcontentloaded"), which returns long before the lazy chat chunk resolves, so those race by construction rather than only on a slow runner. Reproduced on main at aa32c1861: Aria snapshot: - text: Loading....

It shows up on both smokes, not one: Windows fails at line 135 (the first assertion, after the post-change-password route transition) and macOS at line 189 (the second, after a reload). Routes all six through one helper that waits on the same budget the composer already uses.

6. The playwright stub answers only one name

test_heavy_thread_measurement_integrity.py installs a stub into sys.modules for playwright and playwright.sync_api at collection time, defining exactly sync_playwright = None. sys.modules is session-wide, so every later test imports its harness through that stub. Harnesses that need only sync_playwright were satisfied; playwright_strip_ansi_smoke.py imports Page, expect, sync_playwright and so was the only one that broke, which is why just one of four parametrisations of test_an_empty_smoke_base_url_means_unset failed. (unknown location) in the error is the tell for a bare ModuleType, not a namespace package, and importorskip cannot help because the stub is importable by then.

The stub now answers any public name with a callable that raises, so nothing can silently pretend to drive a browser, and dunders still raise AttributeError so it does not masquerade as a package. Both victims keep asserting exactly what they did before; nothing is skipped.

7. The heavy-thread smoke stubs a fork-count URL the app stopped requesting

#8992 replaced the per-message /api/chat/threads/{id}/messages/{id}/forks with one per-thread /api/chat/threads/{id}/forks. The smoke's in-page allowlist was not moved with it, so every fork-count GET reached the dev server and the harness failed its own self-check: HARNESS-BROKEN ... let 2 /api/ requests reach the network during the measured actions. The store refreshes on CHAT_HISTORY_UPDATED_EVENT, which the delete action fires, so two of the three landed inside the measurement. Harness bug, not a frontend one: the app is right to make one request per thread.

The allowlist entry is corrected, the failure message now names the offending URLs, and a new contract test pins the stub against the URL chat-api.ts actually builds so this cannot drift again. The self-check keeps zero tolerance for strays.

Why these two were invisible until now. On every other open PR the Unit tests step in Frontend build + bundle sanity fails on item 1, so Install Chromium for browser smokes and every Browser smoke for * step is skipped, and Repo tests (CPU) reported the 71 failures of items 2 and 3 rather than these. Fixing the import is what let the job reach them. With 6 and 7 fixed, the three smokes that were queued behind the heavy-thread abort (settings tab panels, a settings panel that cannot load, chat stream pacing) were run locally and are green, so nothing further is hiding.

Verification

All test-only except the one-line import fix. 255 pass across the two repo-test files, 189 across the two backend files, and the frontend suite is green again. Both test changes were mutation-checked: reverting the source behaviour they guard makes them fail. My first attempt at 3 was too weak and passed under mutation because cpuFallbackReason also appears in the function signature, so it is now scoped to the description expression.