feat(mining): deterministic, auditable analysis

[tuzuminami]

Summary

Closes #83.

• Introduce one deterministic analysis-preparation pipeline with UTC ordering, source-identity exclusion rules, configurable session boundaries, and an analysis receipt.
• Carry the identical receipt through JSON, Markdown, Mermaid, Draw.io, CSV sidecar, automation candidates, and the manual LLM handoff bundle.
• Make low-confidence case correlation repairable in the desktop UI with a reason-coded local audit record; surface reason-coded exclusions in event quality.
• Reuse immutable analysis snapshots across concurrent local dashboard requests and invalidate them after every event or analysis-setting mutation.
• Add mixed-offset golden data, fixed-seed generated invariant coverage, export equivalence tests, and output calculation contracts.

Why

Previously the product could infer mining results but did not preserve one verifiable input/exclusion/accounting context across every export or offer an auditable correction path for low-confidence cases. Concurrent dashboard routes also rebuilt the same large analysis repeatedly, causing localhost request timeouts.

User impact

Non-engineers can see what was used or excluded, why it happened, and correct a case assignment locally. An external LLM receives reproducible process-mining context suitable for Mermaid generation without implementing any LLM integration. Large local datasets keep the dashboard responsive during concurrent refreshes.

Validation

• ./scripts/test.sh — 100 Python tests, 8 WebUI tests, smoke and lifecycle smoke
• ./scripts/perf_smoke.sh
• ./scripts/perf_http_smoke.sh — 100k concurrent dashboard: 5.612 seconds (15-second limit)
• ./scripts/lint.sh
• ./scripts/check_licenses.sh
• ./scripts/check_no_external_network.sh
• cargo check --manifest-path apps/desktop/src-tauri/Cargo.toml --locked
• git diff --check
• Two read-only acceptance reviews: pass