fix(web): copying commands with prose no longer adds code fences#24433

Problem

Selecting a command in an assistant reply and extending the selection into adjacent prose adds Markdown code fences to the clipboard. Pasting into a terminal includes the generated bash fence and closing fence instead of just the selected command and text.

Change

Copy ordinary code blocks in mixed selections without generated fences. Protect their contents from prose whitespace cleanup, retain Mermaid source whitespace, and prevent internal boundaries from leaking through table Markdown export. Code-only selection, rich HTML clipboard data, and the independent Copy code / Copy message buttons retain their existing behavior.

Scope and approval

Fixes https://github.com/pingdotgg/t3code/issues/23930 GitCafe: https://git.cafe/pingdotgg/t3code/issues/23930

Triaged bug via the automated issue flow. Changes are limited to the web clipboard serializer and its existing tests; desktop shares this renderer. Native mobile uses a separate copy path. No setting, persisted state, contract, or provider change.

Verification

• Reproduced in the real base dev app: selecting printf 'hello\n' through the first T of the following paragraph and using native Copy/Paste produces fenced text. The regression test fails on base for those added fences.
• Focused clipboard tests: 8 pass after integrating current main. Targeted lint passes; web typecheck passed during implementation. Both independent code reviewers returned MERGE after the Mermaid and table findings were fixed.
• Lead integrated browser check: the same selection copies printf 'hello\n' followed by two newlines and T, without generated fences.
• Captures use base a46f80ee28 and implementation 6148f48beb. The later main merge leaves both changed files byte-identical to the reviewed/captured implementation.

Evidence: full index: 72 before/after pairs, diff images, eight recordings, exact clipboard values, and uploaded drivers, fixture matrix. Reproduction-only artifacts are uploaded separately and are not committed.

Recorded interaction: select command through the first T → Control+C → click the unsent plain-text destination → Control+V. The index gives UTC timestamps for every row. Before recording / After recording.

Rendering environment: collaborative Chromium browser, 1440×900 and 390×844 CSS pixels, light/dark, DPR 2, default font scaling. Wide stills export at 1280×800 and narrow stills at 780×1688. Every retained still passed the identical-frames-500-ms-apart check. Diff0 is absolute RGB difference amplified ×8; DiffT retains channel differences greater than 16, amplified ×8.

Visible differences: pasted text loses generated fences for ordinary code in mixed selections. Two narrow inside-diagram pairs also contain a 34-pixel residual in a Python language icon; its cause is unverified. No broad rendering regression was observed.

Proof audit: INSUFFICIENT for the full matrix, while the reported fence-removal behavior is supported. The auditor independently reproduced all 72 diff counts and rebuilt the main narrow after case with matching reply and paste destination. Remaining proof: narrated continuous interaction review; whitespace/selection edge variants; resolved-context MIME; file/attachment preview; null/skipped/multi-range selections; reliable table Markdown/CSV and narrow button controls; and native Electron clipboard delivery on Linux/macOS/Windows. DPR 1 was unavailable. Long pasted outputs are clipped by the evidence textarea. The independent rebuild differed in the lower composer region. Native mobile is outside the changed path. CI and bot review are pending; the bugfix handoff label has not been added.

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