fix(web): collapsed section headers point their chevron down#24438

Problem

With Snoozed or Settled threads present, collapse either sidebar section: its chevron points up. Expanding points it down, opposite to the app's other disclosures.

Change

Rotate the shared down chevron when expanded. The same correction reaches Working, Previous agents, and workflow phase headers; content visibility, labels, and click handling are unchanged.

Scope and approval

Fixes https://github.com/pingdotgg/t3code/issues/23950 GitCafe: #23950

Triaged bug via the automated issue flow. One shared visual condition; no new option, state, contract, or provider behavior.

Verification

Capture caveat: desktop shelf/drag comparisons retain a small unrelated brand-text raster difference near the top left; it is visible in the unmasked diffs. No claim that nothing else changed.

• Reproduced in the real web app on main with isolated seeded threads; both shelf headers were inverted while visibility and aria-expanded were correct.
• After merging current main, vp lint apps/web/src/components/ui/collapsible-section-header.tsx and vp run typecheck in apps/web passed. No automated test was added for this visual-only correction.
• Two independent code reviewers (GPT-6.1 Sol and Claude Opus 5.5) returned MERGE recommendations.
• Web at CSS 1440×900 and 390×844, light/dark; Chrome 154 on Linux, DPR 2, root font 16px, zoom/font scaling 100%. Exported PNGs are 1280×800 and 780×1688 respectively. Clock frozen, fonts awaited, reduced motion and transitions disabled consistently. Each accepted endpoint has an RGB-identical partner at least 500 ms later.
• Desktop uses the shared web component; a separate Electron runtime was not exercised. Native mobile uses a separate, already-correct implementation and is unchanged. No connection, provider, server, persistence, or permission paths changed.
• Captures below are from before a46f80ee28 and fixed 0593c8bc89, before the final main merge. The final diff remains the same single condition; the shared header and its callers were unchanged on the merged main.

Fixture matrix · Combined evidence index, recordings, seeds and drivers · Original recordings and narrated steps

S = all three shelves; R0 = plain Previous agents; R1 = failed Previous agents; W = running/done/failed workflow phases including plural count; D1 = emphasized Settled during drag; D2 = accent Settled drop target. Intended visible differences are only the shared header rotations. Other labels, counts, tone, and content match within each selected pair, except the disclosed desktop brand raster differences.

Evidence

Diff0 is absolute RGB difference amplified ×16, without threshold. DiffT marks maximum-channel differences >16/255. No masks or excluded regions.

Interaction proof

The responsive light shelf recording shows the issue directly: independently open and close Working, Snoozed, and Settled; content visibility agrees with the corrected orientation. The linked original index gives approximate click timestamps and paired recordings for each retained row. The repair index supplies replacement recordings, sampled inspection frames, exact-settling partners, and fixture drivers. Drag recordings hold without dropping, then cancel; collapsed accent uses a selected Settled thread and a slim Snoozed source. Workflow recordings keep the outer toggle open and toggle each phase.

Independent final proof audit: SUFFICIENT for all six visual rows; independently checked 48 combined pairs and 96 exact settling comparisons. Recordings were sampled rather than reviewed continuously. R0 proves the plain header; three retained combinations use an Idle child rather than Done, so completed-status behavior is not claimed there. Capture head was not rebuilt after the final main merge.

Implementation: Claude Opus 5.5 via Claude Agent harness. Lead, reproduction, evidence and review: GPT-6.1 Sol via Codex harness; independent review: Claude Opus 5.5 via Claude Agent harness.