The full-label width check hid current-line attribution when only part of the commit metadata fit. Render into the remaining cells without covering source text; preserve offscreen and no-space guards.
Diff and snapshot sources were excluded before scheduling, and the separate diff renderer never painted attribution. Capture per-side Git provenance, preserve private object stores for historical tabs, and reuse the current-line worker and cache.
Replace the 16 MiB comparison rejection with independent bounded per-side loading and frozen Working blobs. Bound diff computation, defer large inline highlighting, index source-row navigation, and expose actual load failures. Verified the real 52.3 MB parser through repository/session APIs and terminal rendering with EOF/start navigation.
The label lag came from repeatedly projecting full viewports while walking backwards from the caret. Skip only candidates whose maximum possible source/wrap and label height cannot reach the target, then retain the original first-target search. Reuse already resolved scope contexts. The conservative bound preserves non-monotonic unified declaration visibility and independent sticky labels. Added regressions for split declaration-visibility ties and unified height changes; architecture updated. Verification: 10,176 differential projections matched; all 964 release tests pass; release binary built. Reduced repro p95: 3.9572ms -> 1.4467ms. Live 160x100 function-dense 360-key burst median catch-up: 22.895ms -> 9.611ms, with final cursor/viewport asserted. Full and no-dependency Clippy checks are blocked by existing repository warnings, including vendored Crossterm, blame, LSP, Markdown, syntax and existing TUI code. Temporary harnesses removed. No remote changes.
Valdemar Lorenzen
hl-valdemar