test(store): benchmark snapshot persistence and SQLite facade contention

[DeliciousBuding]

Summary

Closes #2333
Refs #2256 (remaining FileStore / SQLite measurements, not EventLog #2304).

• Add FileStore snapshot / indented JSON discard-encoding / actual Sync+rename save / explicit Flush+LastPersistError benchmarks at fixed resident scales.
• Add real SQLiteStore memory-reader / synchronous durable-writer contention workloads without changing MaxOpenConns, PRAGMAs, timers or production source. Track connection-pool wait deltas and SQL total_changes() outside the timed workload.
• Bound instrumentation memory with per-reader latency histograms. Quantiles are clock-observed power-of-two bucket upper bounds; zero clock readings are explicit, not a claimed 1 ns operation. Each reader cycle is one GetRun plus one ListRuns for a ten-run thread.
• Validate complete persisted bytes and restored snapshots outside timing, including content/order/checkpoints; SQLite readback occurs before Close can hide a missing write with another flush.
• Clarify the existing memory-read / durable-write boundary in the Edge architecture owner document.

Validation

• Existing store package tests passed on the starting baseline.
• Full store package tests plus all 21 benchmark cases passed under Linux race instrumentation; final instrumentation/readback changes passed a second all-case race smoke (92.007 s).
• All 21 final benchmark cases passed on Windows with the pure-Go SQLite driver (CGO_ENABLED=0). These are correctness smoke runs, not performance samples.
• go vet ./internal/store, staticcheck 2026.1, verify-doc-ssot.py, and diff check passed.
• Formal non-race Linux measurements passed: 105 samples (21 cases x 5 repetitions), -benchtime=1s -cpu=4 -benchmem. Methodology, all raw samples, medians/ranges and bounded conclusions. All seven required CI gates passed.
• Retain FileStore and single-connection/WAL configuration. The measured unchanged-checkpoint amplification is tracked separately in fix(store): prune run checkpoints and persist their delta state #2335; this benchmark PR does not include its production fix.

Boundaries

Only generated local temp-store fixtures; no user workspace contents, live database or service changes. Fixed resident entity counts and unthrottled synthetic reader loops are not production request rates. Allocated bytes are process-wide (including concurrent readers), not RSS or writer-only heap. The short workload does not exercise the five-minute background maintenance timer. FileStore's explicit flush workload does not measure its debounce frequency. No performance threshold or new CI job.

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Team

Run ID: a454893f-c944-42ac-ad89-434533548841

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[DeliciousBuding]

Final verification at 3e326515b46ade47692012dbcc45e1e563ce8d7b: all seven live required checks (validate, go-edge, go-hub, windows-go, windows-frontend, backend-required, frontend-required) are present and SUCCESS; merge state CLEAN; no review threads/submitted reviews. Automated review skipping is not treated as review approval.

All 105 formal samples and bounded decisions are saved. Local final all-case Linux race and Windows smoke passed; no new timing gate or production change. The concrete checkpoint follow-up is #2335, kept out of this benchmark PR.