Qt-removal R7.6b: the cutover - delete graphlink_app/, burn-down pin reaches 0

[dovvnloading]

The cutover. Deletes graphlink_app/ (289 .py files, 149 of them importing Qt) and the three Qt files left in graphlink_plugins/. Zero files in the repository import PySide6 or PyQt. qt_burndown.json goes 152 → 0 and the gate test's zero-mode assertions now carry the load.

The one hard entanglement

npm run check:schema shelled out to graphlink_app/graphlink_island_codegen.py, so deleting the directory would have broken the frontend build.

Its closure moves to a new top-level contracts/ package — codegen.py, payload_schema.py, and the 11 payload dataclasses the SPA imports. Recon confirmed all 25 payload modules were already Qt-free with zero non-stdlib imports, so the move needed no rewriting.

The registry is pruned 25 → 11 here rather than in #156, where doing it while the dataclasses still lived in graphlink_app/ broke 61 tests. The dataclass and its generated artifact are one unit and had to be deleted together.

Test coverage, stated plainly

The 19 per-payload schema tests died with graphlink_app/. They could not be ported one-for-one — every one imported a Qt bridge class (DragSpeedBridge, NotificationBridge, …) to exercise the payload alongside its publisher.

contracts/tests/test_generated_artifacts.py replaces them by parametrizing over GENERATED_ARTIFACTS, so it covers all 11 surviving payloads where the old files covered 5, and a new registry entry is guarded automatically instead of shipping unguarded until someone writes its file. The codegen CLI's drift tests are ported, repointed from composer-state to scene-state.

Genuinely lost: the bridge-publishes-a-valid-payload round-trip, whose Qt publishers no longer exist.

Backend test count drops 2505 → 961 because ~1550 tests lived in graphlink_app/tests/.

Gate test

The find_spec("PySide6") is None assertion is deleted. It tested whether the developer's environment had a Qt wheel installed, not whether Graphlink contains Qt — a contributor with PySide6 installed for an unrelated project would have failed a gate they hadn't broken, with pip uninstall as the "fix" for a red build. A declaration check across pyproject.toml and requirements* replaces it.

Negative-testing that rewrite found a bug in it: the first version globbed requirements*.txt, which misses requirements.in — the pip-compile source. Qt could have been re-added there and passed the gate until the next recompile. Both probes (a Qt-importing file, a manifest re-declaring Qt) now fail the gate as they should.

Two breaks found while deleting

• [project.gui-scripts] still pointed at graphlink_app:main, which no longer exists. Repointed to graphlink_desktop:main, the pywebview shell.
• The py-modules/package-dir indirection existed only to map graphlink_app/'s contents to the top of the distribution. With the directory gone it collapses to a flat layout — which closes both packaging gaps R7.2 recorded as accepted-for-now: the 17 relocated modules are listable again, and version = { attr = "graphlink_version.APP_VERSION" } now resolves instead of hard-failing a real build.

Test plan

• python -m pytest -q (what CI runs, whole repo): 961 passed
• npm run check: 779 frontend tests, 0 lint errors, typecheck/build clean
• check:schema passes from its new home; 11 artifact sets up to date
• Gate negative-tested in both directions, not just observed green
• qt_burndown.json = 0/0/0; zero Qt-importing files repo-wide

Remaining

R7.6c: README/CONTRIBUTING/PR-template rewrite and the live §0 GATE drive. This PR is the structural end of the Qt removal; R7.6c is documentation and acceptance.