fix(refresh): stop merge-view jitter and preserve session state

[esmuellert]

Summary

Fixes #558.

Stop merge-view jitter and repeated refresh/reinitialization by treating repository notifications as input invalidations, not instructions to reopen the selected comparison. Refresh only what changed, while keeping working buffers and editable merge Results authoritative.

This also consolidates refresh ownership, adds real-Git/Neovim regression coverage, and separates E2E, integration and unit tests.

Refresh model and ownership

• Preserve the worktree / index / head / refs classifications and the 500 ms polling fallback, including native-process failure recovery.
• Keep symbolic source identities separate from resolved revisions. Settle and compare actual input contents before updating a comparison, including unchanged-status worktree edits and mutable index/ref inputs.
• Separate panel-list refresh, comparison refresh and merge input reinitialization. Unrelated changes must not disturb pane identities, focus, cursor, viewport, folds or an unchanged comparison.
• Protect edited Result buffers when merge inputs change: warn rather than silently replacing the user's work. Resolution actions update decorations without reinitializing Result content.
• Reject stale results after selections, buffer replacement, tab closure or disposal; respect shared-buffer ownership during cleanup and restoration.
• Make ui.refresh the single controller. session.panel.data owns list data, query parameters and selection; panel views consume notifications and keep only local presentation state.
• Move Git reads/comparison definitions into refresh/panel.lua and refresh/inputs.lua, and reuse existing view renderers for in-place updates. Remove the old auto_refresh, Explorer scheduler and per-panel refresh modules.

Regressions caught by the expanded tests

• Reappearing files could block on W13 after a deleted-file preview.
• Quoted rename paths, literal -> filenames and renamed index paths could be read incorrectly; status parsing now uses NUL-delimited paths.
• Parentless History commits could reference an invalid parent or omit their files.
• History refresh could lose line-range/reverse options.
• Inline revision updates could trigger unwanted FileType/LSP attachment.
• Following a file outside Git could retain a stale rather than empty comparison side.
• Shared-buffer cleanup and inlay-hint restoration could interfere with another session.

CI follow-up: e6d6678

The first cross-platform run exposed issues that were not visible on the initial fast local machine. These were reproduced and corrected rather than handled by weakening result assertions:

• Slow polling: when a read exceeds 500 ms, repeated fallback ticks must not create an endless backlog. Polls now sample only while idle; real classified invalidations still queue during in-flight reads. Added controller regressions and E2Es delaying real Git results by 700 ms.
• Synthetic gutter fixtures: automatically attached comparison controllers were overwriting hand-authored projections on slower runners. These component fixtures now explicitly retire that unrelated controller after setup; real merge E2Es keep it active. A delayed screen assertion verifies the fixture survives beyond the polling interval.
• Atomic index changes: fixtures now write Git blobs/index entries directly, avoiding accidental intermediate working-tree states during race tests.
• Windows paths: literal-arrow cases use real skip-worktree index entries, so the same Git/parser/UI assertions run on Windows without creating an illegal > filesystem path. The cases are not skipped on Windows.
• Windows cleanup: plugin-aware fixtures retire their own sessions/watchers before deleting directories, without closing other fixtures' sessions.
• Progress and budgets: the supervisor prints spec starts and active-worker heartbeats. Windows runs four workers with a bounded 15-minute whole-spec budget (large specs contain up to 120 isolated Git/Neovim cases); individual asynchronous assertions remain bounded.

Upstream scrollbind distinction

Neovim reverted #41519 in 0c9012f, restoring its known tall-virtual-line oscillation. The previous nvim-0.13 test gate incorrectly assumed the fix was still present. The unchanged strict upstream assertion is now explicitly opt-in via CODEDIFF_TEST_UPSTREAM_SCROLLBIND=1 on a build carrying the fix; it was verified on the earlier fixed build. CodeDiff's own scrollbind setup remains mandatory. No new scrolling workaround or claim to fix Neovim's reverted behavior is included.

Tests and fixtures

• Shared TMP repository/worktree factory: independent refs, indexes and object databases, no hardlink sharing, deterministic reusable profiles, Git environment isolation and cleanup guarantees.
• Plain repositories, unborn branches, bare local remotes and SHA-256 use the same lifecycle.
• Real commands/keys, Git changes, native/polling notifications and Neovim state assertions. Embedded-UI cases additionally assert screen-grid text and colors; race tests delay real results rather than fabricate contents.
• Runnable specs are grouped under tests/unit/ (13 files), tests/integration/ (94 files) and tests/e2e/ (16 files). Support, fixtures and framework implementation are separate.
• The directory reorganization preserved all 1,590 existing expanded registrations and added 5 discovery checks. CI regressions add 10 more, for 1,605 current registrations.
• The refresh change-audit matrix has 133 named scenarios / 548 parameterized E2E executions. The complete E2E directory has 591 executions, including previously scattered command/working-file tests; these are different counts, not 591 newly added scenarios.
• CLI entry points support layers, directories and individual specs; empty or invalid selections fail explicitly. CI paths and documentation are updated.

Validation

PR Validation run 34900369664 completed successfully: 14/14 jobs.

CI platform/check	Result
Linux x64 and ARM64	Build, native tests and all 123 Neovim specs passed
macOS x64 and ARM64	Build, native tests and all 123 Neovim specs passed
Windows x64 and ARM64	Build, native tests and all 123 Neovim specs passed
Android ARM64 and x64	Cross-build and ABI checks passed; these jobs do not run Neovim tests
Standalone builds	All five platform/OpenMP configurations and command smoke E2Es passed
Diff regression gate	Passed

The desktop jobs use official Neovim nightly v0.13.0-dev-1638+ge7e29d9b6d. Windows logs confirm the formerly failing arrow-path cases, slow-read cases, 30 gutter checks and 8 repository-helper checks actually executed and passed.

Additional local validation:

• The same official nightly: all 123 spec files pass with eight concurrent workers.
• Neovim 0.12.3: all 16 E2E specs / 591 executions pass. The full 0.12.3 suite is not claimed green: its pre-existing gutter-coordinate differences remain outside this refresh change.
• make build, C tests 10/10, StyLua on changed paths and git diff --check passed.
• CLI selection, invocation outside the checkout, invalid-target handling, registration preservation and reference/link audits passed.

Review notes and scope

The original seven commits remain intact, followed by the CI correction commit e6d6678; no history was rewritten. The test-organization commit 2dc5a5a changes no production Lua behavior. Most of the large file count is test relocation.

Merge algorithms, filler placement/count rules, gutter implementation, highlight definitions and unrelated scrolling behavior are unchanged.

• Behavior-to-scenario coverage map
• Test organization and execution
• Fixture graph and reproduction

The matrix is behavioral evidence, not an absolute safety or exhaustive branch-coverage guarantee.

[esmuellert]

CI follow-up at e6d6678: PR Validation run 34900369664 has completed successfully (14/14 jobs), including the full 123-spec suite on both Windows architectures. All required checks are green; auto-merge remains enabled and still awaits the required review approval. The separate GitHub-managed Advanced Security AI run 34900367001 failed with a service-side CAPI 400 "The requested model is not supported"; it is not a required check, and no security check was disabled to obtain the Validation result. CodeQL and GitGuardian checks passed.