Fix #613 + #615: enable CI, and make the snapshot fixture hermetic

[JDonaghy]

Fixes #613
Fixes #615

CI is green on this branch (d4682ce) — the first green CI run in this repo, and the first time #615's fix has been validated in the environment that was actually failing.

Why these two are one PR

They deadlocked. ci.yml triggers only on pull_request: branches: [main, develop], so:

• No CI at all: both workflows are checked in as *.disabled, so every vimcode PR is blocked by the merge gate with 'no checks reported' #613 (enable CI) was red — on exactly the six snapshots Six render snapshot tests pass locally and fail in CI: nerd-font glyph widths resolve differently, so the Test stage has been reporting green on machine-dependent assertions #615 fixes.
• Six render snapshot tests pass locally and fail in CI: nerd-font glyph widths resolve differently, so the Test stage has been reporting green on machine-dependent assertions #615's PR (#615: Six render snapshot tests pass locally and fail in CI: nerd-font glyph widths resolve differently, so the Test stage has been reporting green on machine-dependent assertions #616) had no checks at all: it branched before CI existed, so the merge gate refused it with "no checks reported".

Stacking wouldn't help — a PR based on this branch isn't targeting main/develop, so it gets zero checks too. The fix had to ride the branch that enables CI.

The four commits

commit	what
de40cdc	re-enable CI as a minimal headless test gate (#613)
43341b0	clone quadraui as a sibling in CI — actions/checkout can't write outside the workspace (#613)
3a4fc3f	revert the PUA-glyph desensitization — its root-cause theory was false
d4682ce	make test_engine hermetic against the runner's session.json (#615)

The revert-then-fix history is deliberate, not churn: the reverted commits were written against a root cause that turned out to be wrong, and they had already been approved on #616 before that came to light. Reverting them is itself part of the fix.

#615's real root cause

Four hypotheses were investigated and eliminated, three of which had been asserted as verified fact on the issue:

• unpinned unicode-width >= 0.2.0 — a diagnostic step run inside the failing CI job reported LOCK UNCHANGED — cargo honored the committed lock; 0.2.2 on both sides
• rustc toolchain skew — cargo +1.97.1 passes locally, byte-identical
• cargo feature unification (--no-default-features compiling a differently-featured path-dep) — 2180 passed / 0 failed both ways
• quadraui source skew — origin/develop == 7d446ec == every local checkout

The actual cause: test_engine() was not hermetic. Engine::new() calls SessionState::load(), which reads ~/.config/vimcode/session.json from the test runner's $HOME. That file's explorer_visible decides whether the constructor calls app_shell.hide_sidebar() — and the fixture never undid it:

session.explorer_visible → Engine::new() → app_shell.hide_sidebar()
  → build_activity_bar(): is_active = sb_visible && …
  → quadraui_tui.rs:92 — set_cell(…, '\u{258E}', …)     ← the ▎ that goes missing

Every dev and fleet machine has that file with explorer_visible: true from someone actually running vimcode. A GitHub runner's $HOME never has, so the loader falls back to false, the sidebar starts hidden, and six snapshots that assume a visible sidebar fail. Proven by bisection: seeding a fresh $HOME with {"explorer_visible": true} made all six pass.

Note a CI-shaped checkout was not enough to reproduce — shallow clone, detached HEAD, /tmp path, CI=true all passed 2180/0. The variable was the home directory, never the checkout.

This is the fourth leak of this class in that one fixture; the #439 comment block right above the fix documents the first three (git_branch, sc_ahead, sc_behind). The structural fix — an Engine::new_for_test() that never reads disk or git — is filed as #617, deliberately kept off this branch so it doesn't ride the critical path.

Reviewer notes

• The fix is 17 lines in one file, entirely inside #[cfg(test)] mod tests. No production code path changes.
• e.session alone is insufficient and the comment says why: hide_sidebar() has already run inside Engine::new(), so the sidebar is restored through app_shell's existing public sidebar_visible()/toggle_sidebar() API. No new quadraui API, no quadraui files touched.
• No .snap file is modified and cargo insta accept was never run. The snapshots were always correct; the fixture was wrong.
• Verified both ways on dellserver: normal $HOME and HOME=$(mktemp -d) each report 2180 passed; 0 failed. The clean-home run is the one that matters — an ordinary cargo test structurally cannot catch this bug.
• Prior approvals on this PR and on #615: Six render snapshot tests pass locally and fail in CI: nerd-font glyph widths resolve differently, so the Test stage has been reporting green on machine-dependent assertions #616 are void under #1488 (content-changing pushes), and the branch is materially different from what either was approved as.

Also filed from this investigation: #618 — icons::USE_NERD_FONTS is an unguarded process-global AtomicBool read during render. Latent today (no test mutates it), but one careless future test away from this same failure shape under parallel execution.

🤖 Generated with Claude Code