…prefix

pnpm v11 uses 'pnpm_<key>' (not 'PNPM_CONFIG_<KEY>') for config env vars.

The transcript display mode now defaults to 'minimal' instead of
'standard'. Updated all default values across Go config layer,
TypeScript lib, bridge mock, and SettingsPanel normalizer.

…in minimal mode

- Replace turn-level segment logic with step-level grouping: each
  assistant + its tool results is one step. Completed non-final steps
  fold into 'Processed' lines. The currently running step shows items
  normally. The final answer renders as-is.
- In minimal mode, expanded TurnCollapse strips reasoning from
  intermediate assistant messages.
- Standard mode unaffected.
- User messages are included as their own groups in step groups for
  correct rendering in both modes.

Instead of rendering one 'Processed' line per step, consecutive
completed non-final steps are batched into a single TurnCollapse
component. The batch is flushed before rendering the final answer
or an active (streaming) step.

A new assistant message appearing in items marks the end of the
previous step (regardless of streaming state). This ensures
completed steps are folded immediately when the next step begins,
and the last non-streaming assistant is correctly identified as
the final answer.

…next message

When a new user message appears, the previous turn's last assistant
group must be marked as isFinal=true if it's a non-streaming text
response. Previously it was always isFinal=false, causing the final
answer of completed turns to be folded into 'Processed'.

Read-only tools (read_file, grep, ls, etc.) now render during execution
so the user sees progress, then are removed from the DOM after completion
(including errors and failures). This keeps the transcript focused on
writer tools and model responses.

… and compaction cards

Replace ProcessCard-based card rendering with lightweight inline elements:

- ToolCard: replaces ProcessCard with button-based head row + collapsible body,
  removes processState/processTone helpers. Uses simple text status icons.
- Reasoning: collapses thinking block using native useState + ChevronDown toggle
  instead of ProcessCard. Removes ProcessStatusIcon dependency.
- PhaseCard: single inline div with icon + text.
- CompactionCard: button-based toggle with chevron, pure text content.

All cards now have no card background/border/shadow — they blend into the
transcript background, matching Kun's lightweight inline style.

…read vs search counts

- New ReadOnlyBatch component: folds consecutive read-only tools (read_file,
  grep, ls, glob, web_fetch) under a parent line showing 'Read N files ·
  Search N files' counts, using same chevron-fold style as TurnCollapse
- ReadOnlyBatch default open during execution so user sees progress;
  collapses to summary line after review
- ToolCard quiet mode no longer hides completed tools (visibility managed
  by parent batch or step folding)
- Read (read_file, ls) and search (grep, glob, web_fetch) tools are
  counted separately in the batch header
- i18n: tool.readCount, tool.searchCount in en/zh

…animation

- Remove card backgrounds, borders, shadows from .tool and .process-card
- Unify .tool__head, .readonly-batch__head, .turn-collapse__head into .reasoning__head
- ToolCard: restore collapsible chevron, add data-running, disable summarize display
- ToolCard: wrap name+subject in tool__label-group with 70% max-width truncation
- Reasoning arrow: use ChevronRight + rotate 90deg for consistency
- ReadOnlyBatch: reuse reasoning__head/chevron classes
- TurnCollapse: remove chevron icon, reuse reasoning__head
- Add shimmer-sweep animation: diagonal highlight sweeps across running items
- i18n: add tool.readCount/tool.searchCount, remove trailing > from processedDuration
- SettingsPanel: use live displayMode state from getDisplayMode/onDisplayModeChange