tune: lower coalesce/settle step 40 → 10 ms

[yyoyoian-pixel]

Summary

Lower the adaptive coalesce step (client) and straggler settle step (tunnel-node) from 40 ms to 10 ms. Raise the tunnel-node settle max from 500 ms to 1000 ms.

Rationale

The batch pipeline has two phases where data can accumulate:

1. Client coalesce — ops queue up before firing a batch to Apps Script
2. Tunnel-node settle — upstream TCP replies trickle in before the batch response is sent back

The 40 ms step was conservative: it gave each phase a wide window to pack more data per batch, saving Apps Script round-trips. But this window is wasted on downloads — when the client is just waiting for data and has nothing new to send, each 40 ms step is pure dead air before the batch fires.

The fix is asymmetric by design:

• Step: 40 → 10 ms — When there's nothing else to pack, fire almost immediately. The 10 ms still catches ops that land in the same event-loop tick (e.g. a browser opening 6 parallel connections on page load), so we don't degenerate into single-op batches on a burst.

• Max (client): stays at 1000 ms — When both sides do have data (uploads, bursty page loads), the adaptive reset keeps packing: each arriving op resets the 10 ms step timer, so a rapid burst naturally coalesces up to the 1 s cap. This saves quota by packing many ops into fewer round-trips.

• Settle max (tunnel-node): 500 → 1000 ms — More room to pack straggler upstream replies when targets respond at different speeds. One slow CDN shouldn't force a premature flush that wastes a whole Apps Script call for the late reply.

In short: don't wait when there's nothing to wait for; batch aggressively when there is.

Changes

Component	Setting	Old	New
Client	coalesce_step_ms	40 ms	10 ms
Client	coalesce_max_ms	1000 ms	1000 ms (unchanged)
Tunnel-node	STRAGGLER_SETTLE_STEP	40 ms	10 ms
Tunnel-node	STRAGGLER_SETTLE_MAX	500 ms	1000 ms

Backwards compatibility

Users with "coalesce_step_ms": 40 in config.json keep the old behaviour — the compiled default only applies when the field is absent or 0.

Test plan

• Full tunnel: load a heavy page, confirm downloads feel snappier (~30 ms less per batch)
• Upload a large file via full tunnel, check logs for batch sizes (should still coalesce well)
• Long-running session (Telegram, WebSocket), verify no regressions in push latency
• Android: fresh install picks up 10 ms default; existing config with explicit 40 is preserved

🤖 Generated with Claude Code

[therealaleph]

@yyoyoian-pixel — reasoning is sound and I tested locally (179 lib + 33 tunnel-node tests green). The asymmetric design (small step, generous max) is the right framing — fast-fire when nothing else is queued, but adaptive coalesce on bursts. Merging.

Will ship in v1.9.8 with a note that timing-sensitive deployments can override via coalesce_step_ms in config.json (the env var fallback path you preserved). Thanks.

---

_{[reply via Anthropic Claude | reviewed by @therealaleph]}