refactor(subagent): harmonize `formatToolCallShort` (widget) with `formatToolCallFull` (diagnostics)

[samfoy]

Follow-up from #6. Two renderers for the same event shape now live in two places:

• formatToolCallShort in src/subagent.ts — feeds the live widget, tight budget.
• formatToolCallFull in src/subagent-diagnostics.ts — feeds the failure body's activity trail, generous budget (256-char/line, 20 events).

They cover the same set of tools (bash, read, write, edit, grep, find, ls, unknown) with different truncation rules and slightly different field-lookup fallbacks. If a new tool ships or an existing tool's argument shape changes, both sites have to be kept in sync — and only one is covered by the new unit tests.

Goal
Single source of truth for "render a ToolCallEvent for human display", parameterized on budget.

Sketch

interface RenderOptions {
  maxLineChars: number;      // per-line cap
  pathStyle: 'full' | 'collapsed';  // widget wants collapsed, trail wants full
}

function formatToolCall(event: ToolCallEvent, opts: RenderOptions): string

• formatToolCallShort becomes formatToolCall(e, { maxLineChars: 80, pathStyle: 'collapsed' }).
• formatToolCallFull becomes formatToolCall(e, { maxLineChars: 256, pathStyle: 'full' }).

Non-goals

• Changing the widget's visual output. The unified renderer should produce byte-identical output to the current formatToolCallShort when called with the widget's options.
• Consolidating TrackedRun.lastToolCall (subagent.ts) with activityTrail (diagnostics). They're different concerns — live status vs. post-mortem.

Testing
Move the widget's implicit expectations into explicit tests in subagent-diagnostics.test.ts before the refactor, so the byte-for-byte invariant is pinned.

Deferred from #6 as taste rather than blocker; file this so it doesn't drop.