feat(tools): windowed file_read via offset/limit line range (#74)

[jamiesun]

What & why

Closes #74. Adds windowed file_read so the agent can page through large files instead of reading the whole file and getting blunt-truncated at 8000 bytes. This cuts tokens and makes content past the cap reachable (previously impossible). It pairs naturally with grep's line numbers:

grep → hit at line 420   →   file_read {"path":"…","offset":400,"limit":60}

Changes

• tools/file.zig — readLineRange() → LineWindow {text, total_lines, start_line, end_line}. 1-based offset, optional limit (null = to EOF), editor line convention (trailing newline ≠ extra blank line), clamps offset/limit, empty-file & out-of-range return an empty window. Still bounded by the existing 1 MiB whole-file read limit — this saves tokens fed to the model, not file I/O.
• agent.zig — FileReadArgs gains optional offset/limit. A shared fileReadObservation() helper is used by both execTool and the parallel worker (execReadTool), so windowed reads also work inside parallel. Whole-file path (no offset/limit) is byte-for-byte unchanged. Windowed observation reports 第 X-Y 行 / 共 Z 行, or an explicit out-of-range note. Updated the file_read action description.

Behavior

Input	Result
{"path":"f"}	whole file (unchanged), clipped at 8000 bytes
{"path":"f","offset":400,"limit":60}	lines 400–459, labeled 第 400-459 行 / 共 N 行
{"path":"f","offset":99999}	offset … 超出文件总行数 N

Tests

• readLineRange: mid window, read-to-EOF, limit past EOF clamps, out-of-range, no-trailing-newline, empty file.
• fileReadObservation: whole vs windowed vs out-of-range.

Verification

zig build ✅ · zig build test ✅ (all green). First of the token-efficiency group (#70–#77).