Qt-removal R7.2: relocate 33 Qt-free survivor modules out of graphlink_app/

[dovvnloading]

Relocates everything backend/ has a live runtime dependency on — 17 flat modules (api_provider, graphlink_task_config, graphlink_licensing, the chart/chat/artifact agents, graphlink_grid_view_settings, graphlink_navigation_pins, graphlink_prompts, graphlink_version, graphlink_audio, graphlink_model_catalog, graphlink_secrets, graphlink_token_estimator, graphlink_memory) plus the graphlink_plugins/ package — to sit as ordinary siblings of backend/ at the repo root. Their existing bare-name cross-imports (import api_provider, etc.) are unchanged; backend/__init__.py's sys.path.insert hack that used to reach into graphlink_app/ for this is removed entirely.

content_codec.py stays put, gets ported instead

graphlink_session/deserializers.py/serializers.py (Qt-tainted, still live) import it via a normal package-relative import that a physical move would break. Its 6 functions are ported into backend/canvas.py instead — same "reimplement, don't import" precedent chat_library.py/crash_recovery.py already follow — removing the raw importlib.util.spec_from_file_location workaround this replaces.

graphlink_plugins/ moved wholesale, not selectively

An earlier plan tried to leave 3 still-Qt-tainted files behind (graphlink_plugin_context_menu.py, graphlink_plugin_portal.py, web_research/worker.py) to keep the new location Qt-pure. That broke immediately — ModuleNotFoundError on collection — because Python resolves a dotted import like graphlink_plugins.graphlink_plugin_portal against one specific package, whichever directory of that name comes first on sys.path. Two same-named packages can't coexist. So graphlink_plugins/ moved wholesale (all 19 files) and is not 100% Qt-free at its new home — that closes at the R7.6 cutover.

Two required fixes a design/verification pass missed entirely

It only checked backend/'s own consumption of these modules, not graphlink_app/'s own internal consumption of the files being pulled out from under it:

• graphlink_app/tests/conftest.py's sys.path.insert only reached graphlink_app/ itself — now inserts repo root too, or ~30 test files fail collection.
• graphlink_app/graphlink_app.py (the legacy entry point) needed its own sys.path insert before its first graphlink_* import — verified by actually importing the module and watching it fail at the wrong line when placed one line too late.

pyproject.toml

The 17 relocated modules are dropped from py-modules (setuptools' build_py only ever consults package_dir[""] for flat modules — confirmed empirically that per-module overrides are silently ignored — so a real wheel build won't bundle them; a known, accepted, currently-inert gap since nothing in this project's pipeline runs pip install -e . today). packages.find now searches both roots with a package-dir override for graphlink_plugins — verified via a real python -m build --wheel, which now succeeds. The dynamic version resolution does hard-fail a real build now (verified, not guessed) since it can only ever resolve against package_dir[""], which must stay pointed at graphlink_app/ for the ~89 modules still there — documented plainly as a known gap of the same tier.

Verification

2375 pytest — identical count to before (no test lost or duplicated), compileall clean from repo root, qt_burndown.json's pin re-verified unchanged at 152/84/68, 1053 vitest/tsc/eslint/build unaffected (no frontend touched). Both real entry points smoke-tested directly: importing graphlink_app.py as a module resolves every graphlink_* import cleanly, and backend.app.create_app() resolves everything from its new repo-root location with zero sys.path setup.

R7.3 (port the 3 confirmed test-coverage gaps for now-relocated survivor logic into backend/tests/) is next.