test(precompute): add real-pipeline e2e regression test for #474

[milindsrivastava1997]

Summary

• Stacked on test(precompute): add a wall-clock override hook to PrecomputeEngine #537. Adds JsonFileIngestConfig::batch_delay_ms (default 0, a no-op everywhere in production and in the existing netflow_single_second e2e test) — a real, explicit delay after each batch send, so an integration test can force ingest to span real wall-clock time deterministically instead of hoping incidental scheduling overhead is enough.
• New e2e test drives the real JsonFileIngestSource → PrecomputeEngine → sink pipeline with a bulk single-timestamp load spread across 10 batches, using test(precompute): add a wall-clock override hook to PrecomputeEngine #537's wall-clock override (unscaled — an earlier scaled version amplified ordinary scheduling jitter into spurious failures) and generous, seconds-scale margins so the periodic flush timer gets many real chances to land mid-ingest without depending on tight timing.
• Verified red against the pre-fix birth-time semantics (temporarily reverted the fix from fix(precompute): keep wall-clock fallback from force-closing panes still under active ingest #536, ran, confirmed rows were dropped mid-ingest, restored the fix) before counting it as a real regression test.
• Along the way, fixed allowed_lateness_ms in the new test's config: every flush_all() call unconditionally nudges the event-time watermark forward by 1ms (the "boundary advance" that lets an idle stream make progress), independent of the wall-clock fallback. With allowed_lateness_ms=0 and multiple batches spaced across several flush ticks, that 1ms drift alone was enough to mark the next same-timestamp batch "late" via a wholly different code path, unrelated to fix(precompute): bulk single-timestamp ingest silently drops rows when wall-clock fallback closes the window mid-ingest #474. A generous, production-realistic value avoids that.

Test plan

• New e2e test passes; verified red against pre-fix code first.
• Existing e2e_netflow_single_second and e2e_precompute_equivalence still pass.
• Full precompute_engine:: unit suite (81 tests) passes.
• cargo clippy -p query_engine_rust --lib --tests clean.

🤖 Generated with Claude Code