fix(web): collapsed diff files stay folded as patches load#23469

Problem

In a large diff, clicking Collapse all before later file patches load lets added, deleted and pure-renamed files reopen. Reproduced in the Linux web dev client with a 24-file, 1.46 MB uncommitted patch.

Change

Use stable previous/current path pairs for diff file identity across placeholder loading and full/lazy refreshes. Distinguish repeated path pairs by occurrence so same-path deletion and addition blocks retain independent folds. Keep content versioning separate so loaded patches still update. Include deterministic fixture generators for rebuilding the evidence.

Scope and approval

Fixes #17674 GitCafe: #17674 Triaged bug via the automated issue flow. Shared web/desktop diff panel only; no settings, persisted state, contracts or provider changes.

Verification

• Focused identity, fold store, file action and file-tree tests: 33 passed across four files after merging main.
• Patch-order follow-up: an active absolute-path diff.orderFile reproduces reordered hashes; explicit order overrides restore identical original outputs across all variants/scenarios. Python compilation passes.
• Fixture follow-up: 11 generator/scenario outputs preserve baseline, patch bytes and hashes under clean and overridden identity/signing/context settings; 21 focused tests and Python compilation pass.
• Web typecheck: passed, with existing Effect suggestions.
• Primary-agent real-app check after merging main: expand all, collapse all and individual added-file expansion work; other files remain folded.

Final evidence index and provenance.

Evidence

Real app with deterministic committed fixtures. Linux Chromium/T3 Browser, default fonts, DPR2, CSS viewports 1440×900 and 390×844, light/dark. Timing HARNESS holds then releases real file responses to make the lazy-load race repeatable.

Before: 70e5e384da. After: 82f820bb7e; affected production blobs are identical at published head 5bd71de3b8. These are retained captures, not recordings of the final merged application revision.

Captioned sequence: hold four pending requests → Collapse all (~0.99s Before / 0.74s After) → release (~1.16s / 1.02s) → endpoint (~9.05s / 6.04s) → toggle individual file types → Expand all (~20.56s / 16.85s). Offsets are approximate logged marks.

Before captioned recording · After captioned recording

Actual desktop endpoint PNGs are 1280×800 despite CSS metadata; original desktop recordings are 2880×1800, annotated recordings 1440×1030 including caption area. Responsive PNGs are 780×1688. Whole-surface comparisons contain incidental sidebar/composer/text/geometry drift across main revisions and imperfect frozen clocks; these are raw pixel differences, not measured regressions. Intended change: later loaded files remain folded and the global folded state stays selected. No styling changes.

Matched unchanged-region control: six modified-file headers at identical responsive-dark coordinates, no resampling/masking/registration, 0 / 262,656 pixels changed; this zero result applies only to the 684×384 crop.

Before and after unchanged modified-file headers

Pixel report. Supplemental After-only threshold/full-lazy transition recording: C3; same-path independent folds: C4. These are controls with mixed supplement provenance, not claimed matched Before/After comparisons.

Native Electron shell was not launched; web and desktop use the same changed component. Native mobile, providers, server and contracts are unchanged. No Android-specific changes.

Supplemental evidence

Final-head supplemental runtime proof at 68e33d0f7c: capture index, provenance, response hashes and reproduction steps. Independent fixture matrix.

Existing limitation, reproduced on main and this PR: Collapse all snapshots existing blocks. A second same-path addition introduced by switching from lazy to full parsing starts open; the deletion remains folded. Comparative probes and both independent reviews confirm this predates the identity fix. No bulk-collapse policy was added.

Final-head second same-path addition reopened after lazy Collapse all and full refresh

New supplemental environment: Linux T3 collaborative Chromium, CSS 1440×900, DPR2, default fonts, light; screenshots 1280×800 and videos 2880×1800. The index discloses the raster mismatch, timings and excluded setup attempts. Endpoint process/cwd and source-map hashes match both changed final-head production blobs. M1–M4 and failure use real server response bytes; M5 injects a generated checkpoint summary and response, with no backend/provider claim. Existing Before/After pixel diffs remain unchanged; new supplements make no pixel-equality claim. What else changed: no source changes; existing evidence's incidental geometry/text/clock differences retain their original caveats.

Fixture portability findings are fixed in 564d37de80. All recorded fixture hashes remain unchanged, so the retained captures apply to identical production code and data.

Model and harness: GPT-6.1-Sol through Codex; Claude Opus 5.5 through Claude Agent for implementation and independent review.