fix: parse Grok Build usage logs

[juliantanx]

Summary

• add a dedicated stateful Grok Build parser for updates.jsonl
• convert cumulative totalTokens counters into positive per-turn deltas
• extract Grok session, model, timestamp, provider, and encoded workspace metadata
• restrict Grok discovery and status counts to actual usage logs
• version the Grok parser watermark so logs skipped by v1.5.10 are replayed once

Root cause

Grok Build's session directory was registered in discovery, but Aggregator had no parser for the grok tool. The generic JSONL parser also does not understand Grok's JSON-RPC fields or cumulative token semantics. Despite producing no records, the CLI advanced each file's watermark to EOF, preventing a later fixed version from automatically importing historical sessions.

Impact

aiusage parse --tool grok and automatic parsing now import existing and newly appended Grok Build usage. Repeated/decreasing cumulative counters do not double-count tokens, multiple sessions remain isolated, and a second parse is idempotent.

Grok currently exposes no stable input/output/cache split, so positive total-token deltas are recorded as input tokens.

Validation

• pnpm test — 605 tests passed across core, web, widget, site, and CLI
• pnpm build — core, web, CLI, and widget builds passed
• focused Grok parser, discovery, watermark migration, and CLI integration tests included

Closes #43

[willtse]

本机测试了下：

结论

项	结果	说明
分支	fix/issue-43-grok-parser (72285c6)	PR #45
Core build	通过	@aiusage/core
单测 grok + aggregator	通过	12/12
集成测 parse-grok / discovery / watermark	通过	32/32
本机 Grok 日志	13 个 updates.jsonl	~\.grok\sessions
首次 parse --tool grok	通过	导入 38 条记录
再次 parse	通过（幂等）	行数仍 38，活跃 turn 可 upsert 更新
会话数	通过	13，与日志文件数一致
模型 / provider	通过	grok-4.5 / xai
cwd 解码	通过	如 D:\git\github_clone
token 语义	通过	增量记入 input_tokens，output=0，无负值
水位迁移	通过	grokParserVersion: 1，13 个文件水位
全量 pnpm test / pnpm build	未跑	非本修复阻塞项
总评	通过	可合入，修复 issue #43

可以认为 PR #45 在本机已验证通过，能合入解决 issue #43。