fix: parse Claude session logs incrementally instead of re-reading whole files

[notcadenwhitt-stack]

Summary

Fixes #337.

On an account with a large Claude Code history (~2GB of session logs this month, single files up to 274MB) the cold scan pegged a core for over a minute without finishing, and because Blob is keyed on whole-file mtime+size, the active session file was re-parsed from byte zero on every refresh.

Three changes, all in the Claude local-log path:

• ISO8601Parser reuses one lock-guarded ISO8601DateFormatter pair instead of allocating per call. This is the per-line timestamp parser for every session log, and was the dominant cost in sample. The lock is needed because the provider caches call it concurrently.
• LocalUsageReader.parseClaudeFile scans lines at the byte level via Data/FileHandle instead of String(contentsOf:) + split, so a 274MB file is never materialized as a Swift String. A new parseClaudeFile(_:fromOffset:fmt:) reads from a byte offset and reports how far it consumed (stopping at the last complete line, so a half-written record is retried next refresh).
• LocalUsageCache.Blob gains an optional offset. In collect, when a cached file has strictly grown, only the appended tail is parsed and merged (dedup by id). Same-size rewrites and truncations still take the full-reparse path, so the existing "mtime changed → reparse" contract holds. Older cache files without the key decode fine (decodeIfPresent) and fall back to a full parse once.

Only the Claude reader opts into the incremental path; every other collect caller is unchanged.

Measured on the history above, release build:

	Before	After
Cold start	60s+ at ~100% CPU, never completed, Claude cache never written	~15s, then idle; cache persists
Peak memory	900MB+ and climbing	~75MB
Warm relaunch / steady-state refresh (active session growing)	full re-parse every refresh	0.0% CPU

Type of change

• Bug fix
• New feature
• Refactor / cleanup
• Documentation
• Other:

UI changes

None.

Checklist

• swift build and swift test pass locally (1165/1166; the one failure, SessionKeySettingsRenderingTests, fails identically on unmodified main in my environment and is unrelated to this change)
• PR title and description are written in English
• UI changes are described above (before/after — images optional)
• No copyrighted assets, secrets, or private tooling references are committed
• Tests were added or updated for this change — four new cases in LocalUsageCacheTests: tail-only parse on append (proven by rewriting the prefix in place and asserting it is not re-read), full reparse on truncation, half-written trailing line retried on the next refresh, and pre-offset cache files still decoding

🤖 Generated with Claude Code