test: add long-running telemetry retention soak

[Korupt-virus]

Summary

• add isolated parser/replay workers that repeat full 20,000-frame lifecycles, force full GC, and record current RSS plus managed heap after every measured cycle
• add deterministic median and ordinary-least-squares retention analysis with focused synthetic contract tests
• run soak weekly on Monday at 08:17 UTC and by manual dispatch; always retain report artifact for 14 days

Detection model

One-shot peak budgets in #259 catch capacity regressions. This soak catches objects retained across repeated parse and replay lifecycle churn.

Ten warm-up cycles are discarded. Default run records 100 post-GC samples per scenario. A metric fails only when both first-to-last ten-sample median growth and positive slope exceed its limits, which separates bounded caches/JIT/allocator behavior from sustained retention:

• managed heap: 64 MiB retained growth and 512 KiB/iteration slope
• RSS: 128 MiB retained growth and 1 MiB/iteration slope

Stack

Temporarily targets throughput-benchmarks while #259 remains open, so this PR shows only soak changes. Retarget to main after #259 merges; do not merge this stacked PR first.

Verification

• bun test test/tooling/replay-parser-soak.test.ts
• bun run typecheck
• bun run lint
• bun run bench:telemetry:soak --iterations=20 --warmup=5 --output=telemetry-soak-results.json
• bun run bench:telemetry:soak --output=telemetry-soak-results.json
• bun run bench:telemetry:current --no-enforce --output=telemetry-benchmark-results.json