Manual verification: macOS-driven harness changes on a Linux/GNOME checkout

[jonocodes]

What to verify

A macOS upkeep pass (2026-08-06) fixed a set of dev-harness assumptions that only broke off-Linux. Every change was verified on macOS 15.6.1 / Apple Silicon — full ladder green, just test-all 7/7 — but none of it has run on a Linux/GNOME checkout, and all of it touches the shared harness rather than the macOS backend.

The risk isn't subtle behaviour: it's that a Linux dev's just setup or just test-all stops working.

What changed

change	file	Linux risk
uv venv skipped when $VIRTUAL_ENV is set	Justfile (setup-linux, setup-macos)	a Linux dev without flox still gets ./.venv; a Linux dev with an active venv now installs into it instead of creating ./.venv — intended, but changes where packages land
install_evdev_source.sh gets PYTHON="${VIRTUAL_ENV:-.venv}/bin/python"	Justfile	the aarch64 source-build path is Linux-only and was not exercised at all
./.venv/bin prepended to PATH only when it exists	Justfile (test-all)	should be a no-op on a normal Linux checkout
pyright --pythonpath "$(command -v python)"	Justfile (test-all)	with ./.venv present, [tool.pyright] venv still wins — expected to behave exactly as before, but unconfirmed
pyright added to the [dev] extra	pyproject.toml	first just setup-linux after this pulls a new dep (wraps node)
# pyright: ignore[reportMissingImports] on the evdev import	daemon/deckd/input.py	with evdev installed, confirm the ignore doesn't itself become a warning
e2e daemon runs with DECKD_FAKE_INPUT=1	client/playwright.config.ts, daemon/deckd/__main__.py (_build_sinks)	the Linux suite previously relied on the PYTHONPATH=scripts/no-evdev shadow; both are now in play and the shadow must stay redundant, not conflicting
e2e daemon runs with --no-focus	client/playwright.config.ts	on GNOME the daemon previously watched real desktop focus during e2e; specs that implicitly depended on that would now fail
chromium lookup falls back to Playwright's managed browser	client/e2e/find-chromium.mjs	the nix-store path must still win on nix — it's checked first, but confirm no accidental fallback
e2e daemon binary falls back to deckd on PATH	client/playwright.config.ts	.venv/bin/deckd must still be preferred where it exists

Acceptance criteria

On a Linux/GNOME checkout (nix/flox or plain, note which):

• just setup completes and lands packages where you expect (which deckd, python -c "import evdev" where the extra applies).
• just test-all passes all 7 steps.
• Step 1 (pyright) resolves third-party imports — no reportMissingImports storm — and the evdev ignore comment produces no new warning on a box where evdev is installed.
• npm run test:e2e passes, and client/e2e/.daemon.log shows DECKD_FAKE_INPUT set; input is logged, not injected — i.e. no keystroke reached your desktop during the run.
• On nix: e2e still launches the /nix/store/*-playwright-chromium/*/chrome binary, not a downloaded one.
• just dev / just dev-daemon-lan still work normally (the sink and focus changes must not leak into non-test runs — DECKD_FAKE_INPUT unset means real injection).
• aarch64 Linux only, if you have one handy: just setup-linux still source-builds python-evdev via scripts/install_evdev_source.sh.

Notes

Nothing here changes daemon behaviour at runtime; _build_sinks() is a pure extraction of the previous inline sink selection plus the new DECKD_FAKE_INPUT short-circuit (covered by tests in tests/test_input.py). If a step fails, the fix is likely a one-liner in the Justfile — file it as a follow-up rather than reverting the set.