Reduce refresh CPU and memory use

[thrr87]

Summary

• Tail rollout and fact files in bounded chunks. An idle refresh reads no old bytes, and a small append only handles the new records.
• Reuse stable local aggregates and indexes. Release old fact buffers after durable storage, and load receipt detail only when the user opens one task tree.
• Run usage evaluation away from MainActor. Publish only the newest result, including restored state at launch.
• Bound Codex RPC lines and large local reads. Handle timeout races, invalid dates, and integer overflow without terminating the app.
• Cap chart work and reduce needless sync writes without changing analytics meaning or history retention.

Validation

• Swift tests: 495 passed
• Codex client tests with Thread Sanitizer: 43 passed
• Release build: passed
• Full-screen QA: launch, Graphs, Facts, Insights, and repeated manual refresh passed

Representative local measurements:

• persisted restore: 20,019 facts, at most 423 ms per bounded refresh, about 18 MB resident growth
• 1,000 idle refreshes: 82.8 ms total, 0 bytes and 0 records reread, 0 resident growth
• one-record append: 0.287 ms
• stable evaluation over 100,000 facts: 0.182 ms
• 50,000 timestamp parses: 48.8 ms

On the same large local profile, memory grew on each refresh before the retained-buffer fix: 355.4 MB → 480.3 MB → 614.0 MB. After the fix, it settled at 296.8 MB → 472.6 MB → 462.6 MB. A final fresh-build QA run also fell from 641.5 MB to 592.3 MB after another refresh instead of growing again.

Notes

• Large local histories can still raise the process high-water mark while open. Repeated refreshes no longer retain another full fact buffer.
• Controlled source gaps lower coverage or return a clear error. They do not delete analytics history or change its meaning.

Closes #35