fix: stream JSONL reads so token widgets work on large transcripts

[LZong-tw]

Summary

Fixes #550.

Built-in tokens-input / tokens-output / tokens-total widgets showed In: 0 Out: 0 Total: 0 for long-lived Claude Code sessions once the session transcript JSONL exceeded Node's maximum string length (~512MB / 0x1fffffe8).

Root cause

readJsonlLines did:

const content = await readFile(filePath, 'utf-8');

On an ~844MB transcript this throws:

Error: Cannot create a string longer than 0x1fffffe8 characters

getTokenMetrics caught the error and returned all zeros. Token widgets prefer context.tokenMetrics even when zeroed, so they never fell back to context_window totals. Session cost still worked because it comes from statusline stdin, not the transcript scan.

Changes

• Replace whole-file readFile / readFileSync with streaming line iterators:
  • async: fs.createReadStream + readline
  • sync: chunked readSync with Buffer leftover handling (UTF-8 safe across chunk boundaries)
• readJsonlLines / readJsonlLinesSync now assemble lines from those iterators (fixes speed metrics, compaction, session clock, etc. for the same class of failure)
• getTokenMetrics and getSessionDuration iterate lines without materializing the full file as one string
• Tests for streaming/CRLF/empty lines and multi-line aggregation; existing token-metric tests still pass

Verification

npx vitest run src/utils/__tests__/jsonl-lines.test.ts \
  src/utils/__tests__/jsonl-metrics.test.ts \
  src/widgets/__tests__/TokensWidgets.test.ts \
  src/utils/__tests__/jsonl-blocks.test.ts
# 41 passed

Against a real ~844MB production transcript (the session that reproduced #550):

	Before	After (this branch)
getTokenMetrics	all zeros (string-length throw)	totalTokens: 17_794_272_870 in ~7s
In/Out/Total widgets	0 / 0 / 0	non-zero

Test plan

• Unit tests for streaming JSONL reader
• Existing token/speed metric tests
• Real multi-hundred-MB transcript no longer returns zeros
• Reviewer: smoke statusline against a large session if available