Fix fresh-agent tool result attribution and adjacent tool grouping

[danshapiro]

Problem

Claude Code skill payloads inject as synthetic user turns in the fresh-agent transcript, appearing as fake "You" messages. Additionally, adjacent tool-use/result exchanges render as separate entries instead of being grouped into the preceding assistant turn's activity strip.

What the fix does

Server-side (normalize.ts):

• normalizeClaudeTurn now returns FreshAgentNormalizedTurn | null — user turns whose content is entirely system context (e.g. Claude skill instruction payloads starting with "Base directory for this skill:") are filtered out
• normalizeClaudeItem extracts user-authored text via extractUserAuthoredText, stripping system context from user text blocks
• isSyntheticUserTimelineItem filters synthetic user items from timeline pages
• latestTurnId is derived from the last non-null turn instead of the raw resolution value
• normalizeClaudeTurnBody returns null for synthetic turns

Client-side (FreshAgentTranscript.tsx):

• isSyntheticToolResultTurn identifies user-role turns that are entirely tool results
• coalesceSyntheticToolResultTurns folds synthetic tool result turns into the preceding assistant turn via appendTurnItems, or renders them as tool-role turns when no preceding assistant turn exists
• Applied before filterTurnsForDisplay in the display pipeline

utils.ts:

• isSystemContext now detects Claude Code skill payloads starting with "Base directory for this skill:"

Rebase notes

Reconciled with PR #452 (transcript contract) which rewrote normalizeClaudeTurn and FreshAgentTranscript.tsx:

• Old coalesceActivityOnlyTurns/mergeActivityOnlyTurns removed on main; synthetic tool result coalescing reimplemented as a standalone coalesceSyntheticToolResultTurns step in the new displayTurns pipeline
• Type renames adapted: AgentTimelineTurn → ClaudeFreshAgentHistoryTurn, AgentTimelinePage → ClaudeFreshAgentHistoryPage
• Tests updated for main's per-turn activity strip rendering (multiple strips instead of one coalesced strip)

Tests

• 6 new test cases: 2 transcript client tests, 1 normalize server test, 2 utils tests, 1 existing normalize test preserved
• All 50 FreshAgentTranscript tests pass, all 40 server tests pass
• Typecheck and lint clean

Closes kata jb5w.