fix(pretty): coerce string offset and limit in readCall line range

[iamxeph]

Problem

When an LLM provides tool call arguments with string-typed numbers (e.g. offset: "1", limit: "2000", as observed with models like glm-5.3-flash), the read tool summary displays bizarre, inflated line ranges such as:

• lines 1–11999 (for offset: "1", limit: "2000")
• lines 1046–1046399 (for offset: "1046", limit: "400")

Root Cause

In src/summary.ts, readCall calculates the end of the line range as follows:

const range =
  args.offset || args.limit
    ? theme.fg("dim", ` · lines ${args.offset ?? 1}${args.limit ? `–${(args.offset ?? 1) + args.limit - 1}` : "+"}`)
    : "";

While args is typed as { offset?: number; limit?: number }, TypeScript types are erased at runtime. Tool call arguments deserialized from LLM responses may carry string values.
In JavaScript:

1. (args.offset ?? 1) evaluates to "1" (string).
2. ("1") + args.limit evaluates to "1" + "2000" = "12000" (string concatenation due to + associativity).
3. "12000" - 1 coerces to a number via subtraction, resulting in 11999.

Solution

Explicitly coerce args.offset and args.limit to numbers using Number() before calculating the range:

const offset = args.offset ? Number(args.offset) : undefined;
const limit = args.limit ? Number(args.limit) : undefined;
const range =
  offset || limit
    ? theme.fg("dim", ` · lines ${offset ?? 1}${limit ? `–${(offset ?? 1) + limit - 1}` : "+"}`)
    : "";

This prevents string concatenation while preserving the exact original logic and public type definition.

Verification

• Added regression tests in test/pretty.test.ts verifying string-typed arguments for offset and limit.
• Ran tsc --noEmit: clean (0 errors).
• Ran node --test test/pretty.test.ts: all 32 tests passed (0 failed).