fix(web): minimap preview no longer blocks selecting message text#23462

Problem

With Chat width set to Wide, hovering the message minimap opened a preview whose expanded pointer target covered response text. Moving onto that text and double-clicking selected the preview instead of the response underneath.

Change

Keep the minimap's pointer target within the measured navigation strip and make the preview ignore pointer events. Leaving the strip dismisses the preview, so response text remains selectable and links remain clickable. Preview text itself is no longer selectable.

Scope and approval

Fixes #16447 GitCafe: https://git.cafe/pingdotgg/t3code/issues/16447 Triaged bug via the automated issue flow. Reproduction and scope.

Verification

Fixture safety follow-up: 49e0abba89 requires a database-adjacent issue-16447-disposable marker before opening SQLite for writing. Disposable-database checks confirmed missing/invalid markers refuse without changing database bytes or rows; a valid marker seeded successfully. Python compilation and formatting passed. All browser captures below remain from 366ec74; this follow-up changes only the fixture guard and its instructions, with identical production render code. Existing proof-audit limitations remain.

• cd apps/web && ../../node_modules/.bin/vp test run --project unit src/components/chat/MessagesTimeline.logic.test.ts: 81 passed; existing timeline regression checks, not proof of minimap pointer behavior.
• Targeted lint on TimelineMinimap.tsx and MessagesTimeline.logic.ts: passed.
• node_modules/.bin/effect-tsgo diagnostics --project apps/web/tsconfig.json --format text: 0 errors, 0 warnings; 17 existing suggestions.
• git diff --check origin/main...HEAD: passed.

Evidence

Primary Wide recordings reproduce the bug on the merge-base and show message-owned selection and successful underlying link clicks on the fixed head in both themes. The lead also repeated Wide selection and strip navigation in a separate 1280×800 CSS-viewport real-app check. This is capture proof; no implementation-shaped test was added.

Full evidence index (all timestamps, artifacts, and disclosed gaps): issue16447-evidence.md

Rendering environment and fixture

T3 collaborative Browser panel via preview tools only; Chrome 154.0.0.0, Linux x86_64, fine pointer=true, coarse=false, DPR=2. Desktop viewport: measured 1440×900 CSS pixels; video frames 2880×1800. Saved desktop snapshot PNGs are preview-tool-downscaled to 1280×800; this does not mean the CSS viewport was 1280×800. Video-derived primary keyframes are explicitly scaled to 1440×900. Narrow viewport: 390×844 CSS, saved PNG 780×1688. Interface root font 16px; browser scale not changed, text scale 100%. Reduced motion was not emulated. Theme was changed through the normal persisted appearance-mode preference and storage notification; chat width through updateClientSettings, the same settings update path used by the UI. No VITE origin overrides were set.

Fixed committed fixture: issue-16447.py, issue-16447.json. Synthetic-only worktree .t3 projections, retained across serialized server/checkout switches; copied identical scripts into the temporary base checkout without committing. Same origin/environment and fixture data on both sides. Long thread has 80 user turns, assistant prose, links, and reasoning rows; all 80 loaded for desktop long captures. Narrow control loads 10 of the same 80 turns, sufficient for the ≥2-item path. No live userdata writes or requested provider turns. Startup logged zero resumed/requeued provider work. The fixture is a UI projection seed tied to this SHA, not proof of backend/event behavior; rebuilding projections would replace it.

Evidence matrix

Interaction recordings

Script: hover first/middle/last minimap marks, move onto covered response text, double-click, click the reference link, wheel, focus/blur, drag across the rail, then keyboard and intentional rail navigation. Light: selection at 5.8s Before / 5.2s After; link result at 17.0s / 7.1s. Dark: selection action at 4.3s / 4.7s and link result at 5.6s / 6.0s (Before / After). Exact steps/timestamps for every recording are in the linked full evidence index.

Desktop rows below use the measured 1440×900 viewport in their named light/dark theme; M rows use 390×844. Saved snapshots are downscaled as described above. N/Z/C mismatches are unverified regression evidence, not clean comparisons.

Paired keyframes and pixel differences

Zero threshold: absolute RGB difference, amplified ×8 and clipped at 255; every nonzero channel counts. Stated threshold: maximum absolute RGB channel delta >16/255, magenta changed pixels on black, no spatial tolerance. No alignment, masking, crop, or content replacement. Saved snapshots are compared at their original tool-delivered size. Video-derived frames use the same 1440×900 scaling on both sides; lossy video compression contributes to differences. Thresholding does not prove interaction behavior. N/Z and scroll-changing frames deliberately retain mismatches and must not be described as invariant. light-W-video-* rows are inspected video-derived semantic states; all other rows are original snapshot calls and may lag the completed action.

What else changed visibly

Observed intended differences: the preview stops retaining pointer ownership in its former expanded footprint; preview text selection is replaced by the underlying paragraph selection; the underlying link receives the click; pointer exit clears the hover preview; keyboard focus ring is restricted to the actual strip rather than the old expanded 22rem region. Preview position, title/body, and fisheye are visible in the hover recordings. Snapshot calls sometimes show the previous compositor state, so a missing preview in an isolated snapshot is not itself treated as a render regression. Video-derived primary states distinguish that capture lag from the behavior being tested.

Other visible variation retained in the artifacts: selection highlight and selected word, hover/focus bands and ring state, scroll offset and visible message following wheel/drag/keyboard navigation, disabled endpoint controls sampled during navigation, and the N/Z supplement label/padding/content wrap. Earlier recovery captures show a provider update/status toast and open details panel; these were dismissed/closed before the delivered primary paragraph-41 set. Sidebar age/status metadata and transient render/animation/compression differences can contribute to diffs. No unrelated implementation change was made. Clean primary paragraph-41 layout can be judged in the initial Wide states, but the evidence does not establish zero unrelated pixel change across every interaction frame.

Independent audit

Both final code reviewers returned MERGE with no production defects. The independent proof audit verified all 317 local artifacts and exact arithmetic for all 67 diff pairs, and found the primary Wide selection/link proof supported. Its verdict is INSUFFICIENT for D, N20/N8, Z, C2, M-coarse and independent recapture. Those checks remain unverified; this PR does not claim exhaustive visual sign-off.

Proof audit pass2. issue16447-matrix.md; issue16447-lead-browser-proof.json

Models: Claude Opus 5.5 and GPT-6.1 Sol. Harnesses: Claude Agent and Codex in T3 Code.

Takeover

No code changes; evidence and verification added below. The original description, author credit, fixture-safety follow-up and valid primary Wide evidence are preserved. Capture-only work used isolated synthetic state and changed no branch or implementation.

Evidence — recovery 2026-10-22

Source-confirmed matrix · 12 recordings / 23 paired keyframes, steps, both diff thresholds and reproducible scripts · Complete capture package · Independent pass 2 · Current-head recapture and final audit.

New interaction pairs compare BEFORE 3b95e7831b with AFTER 366ec74cf8. The subsequent fixture-guard-only 49e0abba89 has identical production render code and its own independent Wide still below. Earlier recordings retain their original revision labels and are not current-head recordings. Final evidence verdict: INSUFFICIENT.

Current-head Wide light: Chrome154/Linux, no device emulation, fine=true/coarse=false, DPR2, text scale100%, root16px, fonts loaded, CSS1440×900; snapshot1280×800. Loaded80; sidebar256; details closed; g48/h36. Against delivered reference:50129 pixels differ at >0 and20785 at strict max RGB>16, of1024000. Against r5 independent image:91/61, entirely sidebar age text. All comparison images, counts and setup.

Independent current-head AFTER — Wide light paragraph41 preview