perf(refresh): memoize the two steps that dominated every pass

[Gnonymous]

Where the pass actually went

With the animation fixed in v1.1.4, the 30-second refresh became the only CPU event left. Profiling one pass (3179 main-thread samples, ~2 s):

Step	Time	Share	Location
GitCorrelator.buildRanges	951 ms	30%	GitCorrelator.swift:99
Scanner.init → plist decode	745 ms	23%	Scanner.swift:119
CodexScanner.scan	93 ms	3%

Log parsing was not on the list. The mtime+size cache already skips unchanged files, so the incremental-parse idea this task started from would have optimised something that costs nothing. These two steps are the pass.

Fix

Scanner — cache state moves from per-instance to process-wide. Aggregator.run() builds a fresh Scanner every pass, and the on-disk cache is 10 MB of binary plist here, so it was decoded every 30 seconds to reconstruct a dictionary already sitting in memory. Now read once per process, still written back on change so a cold launch starts warm. A lock spans each scan() — passes are already serialized by UsageStore's coalescing guard, so it is uncontended and only makes overlapping passes wait instead of race.

GitCorrelator — buildRanges memoized for 120 s, keyed on the candidate cwds and on the day being reported. The outputs bucket against midnight, so an entry computed yesterday would be wrong today even inside the TTL. Thirty days of git log --numstat per repo cannot meaningfully change between 30-second passes.

Trade-off: a new commit can take up to two minutes to appear in the panel. That is the only visible change.

Measured

Release build, same machine and workload, CPU-time delta (not ps pcpu, which is a decaying average and reads 0 between spikes):

	CPU
v1.1.3	~60% — animation, sustained
v1.1.4	18.7% — animation fixed, refresh pass exposed
v1.1.5	3.2%

60 s window: 1.94 s consumed. 120 s window: 3.85 s — same rate, so it holds across a git recompute.

Tests

Both wins are silent if they regress — the numbers stay correct, the app just burns a core again. PerfCounters exposes the two counters so --self-test can assert reuse (the CLT-only self-test cannot reach the internal types, and XCTest needs a full Xcode toolchain).

• --self-test: 27/27 pass, including 3 new assertions
• Red-green was actually run on all three: they fail with the counters added but the memoization absent, pass with it
• XCTest additions cover the same ground with a real git repo (CI runs these; not runnable locally without Xcode)
• Verified live: scan cache still rewrites within seconds of a log append, git today +607 −56 (10 commits) reflects commits made minutes earlier

🤖 Generated with Claude Code