Scale processing beyond the in-memory session limit

[zrack]

Operator outcome

Large telemetry sessions remain responsive, cancellable, and deterministic under documented memory and latency budgets.

Scope

• Move validation, decoding, aggregation, and bundle construction off the UI thread with Web Workers.
• Add streaming JSON/NDJSON ingestion or a chunked container beyond the current 32 MiB UTF-8 .nlsession limit.
• Evaluate IndexedDB or another local indexed store for frame and metric access.
• Bound downsampling, memory growth, marker lookup, and evidence generation for multi-million-record sessions.
• Add cancellation, progress reporting, and deterministic failure recovery.

Invariants

• Integer-microsecond offsets and half-open ranges remain exact.
• Raw records remain immutable and failure diagnostics remain visible.
• Worker and streaming paths produce the same decoded and evidence results as the current in-memory pipeline.

Exit criteria

• A documented large fixture imports, replays, seeks, selects incidents, and exports without blocking the main UI.
• Memory and latency budgets are measured in automated tests.
• Cancellation cannot produce a falsely complete session or bundle.

[zrack]

Delivered through #27, #28, and #34 for the documented support tier: worker-backed validation, decoding, aggregation, comparison, canonicalization, and bundle construction are cancellable and deterministic; the automated gate imports, persists, reopens, compares, exports, and verifies 200,000 records at 52,378,445 bytes under explicit responsiveness and Chromium heap budgets. This does not claim multi-million-record or out-of-core support.