🤖 perf: sibling subscriptions stall during 1M-row replays; sub-3 s sync blocks left in the replay path

[ThomasK33]

Split out of #4643 (fixed by #4647, which removed the ~10 s token-stats stall). What remains on 1M+ row chats is pre-existing and not an event-loop freeze of that size.

Observed

• activity and terminal activity subscriptions log stalled (no events for ~10-12 s) during the load phase of a 1.24M-row chat, on base and on 🤖 perf: yield while counting tokens so 1M-row chats don't stall the server #4647 (remote UAT: ~11 s, ~32 s after the click, while the server answered pings within 370 ms). Likely cause: the replayed-row flood on the shared WebSocket delays their heartbeats (unproven).
• Remaining synchronous blocks in the onChat replay path (🤖 perf: stream onChat replay with keepalives so slow replays are not aborted #4644): computePriorHistoryFingerprint at the end of a replay (1.4–2.7 s at 1.24M rows) and readTail's buffer.toString().split("\n") over the whole tail (~0.5 s).

Evidence: #4643 root-cause comment and the #4647 UAT chat.

---

Generated with xum • Model: anthropic:claude-opus-5-5 • Thinking: high • Cost: $9.90

[linear-code]

MUX-254

[ThomasK33]

History-read profile from #4653 (1.24M-row single epoch, 500 MiB)

While planning #4653 I timed tokenizer.calculateStats on a fresh xum server (32 cores, load 10–17). The RPC is 54–56 s on main. 18.4–18.5 s of it is HistoryService.getHistoryFromLatestBoundary, and the #4653 PRs do not touch that. After them the RPC is ~26 s, and the history read is ~70% of what remains. A browser open of the same chat ran this read three times (onChat replay, a second read, then calculateStats), at 20–23 s each.

Server CPU profile of one run (--cpu-prof, main isolate, inclusive):

frame	CPU
findProviderHistoryStart (reverse locator scan, no boundary so it walks to byte 0)	12.3 s
— classifyHistoryScanRow (JSON.parse, JSON.stringify(raw).includes(RESET_NEEDLE), zod parts check, normalize)	5.2 s
— reset probes: addHistoryResetProbe 2.3 s self, decodeResetEscapes 1.1 s, escaped-key regex 0.7 s, stripRawResetSeparators 0.7 s	~4.8 s
readTail (forward parse of the whole epoch; one string + split)	2.8 s
findLastBoundaryByteOffset (first read per process, rotation check)	1.1 s
GC (whole profiled run)	4.5 s

So every row is parsed twice, stringified once and zod-checked twice, and the escaped reset probes run on every row.

Idea: a conservative raw prefilter in the locator

A row can only be a boundary or reset if its raw text contains a boundary or reset needle, in plain or JSON-escaped form. If the locator ran a cheap raw check first (for example, no needle and no \u/\x escape anywhere in the row means it can be skipped), it could skip JSON.parse, stringify, zod and the escape decoding for almost every row. That would bring the locator from ~12 s towards a byte scan, the read to roughly 5–6 s, and the calculateStats RPC to ~13–14 s. It would also speed up chat open, which runs the same read.

Boundary and reset detection is correctness- and safety-relevant (the escaped-marker defenses), so it needs its own plan and a differential test (same start offset and rows with and without the prefilter, over fixtures that contain plain, escaped and oversized markers). I have not started it; it is out of #4653's scope by the perf owner's decision.

Harness: /home/coder/perf-runs/4653 on the perf-owner host (e2e/e2e-run.sh with the timing.cjs phase preload; cpuprof.py).

---

Generated with xum • Model: anthropic:claude-opus-5-5 • Thinking: high • Cost: $29.40

[ThomasK33]

Picked up by the perf owner (plan stage).

Scope: the whole-epoch history read (locator prefilter and repeated reads per open), the remaining sync blocks in the onChat replay (computePriorHistoryFingerprint, readTail split), and the sibling-subscription stall. I am measuring on main 5bd46a2 with the #4653 fixtures (100k, 500k and 1.24M rows) and a few large real sessions. The plan (small independent PRs with before/after targets) is in review before implementation starts.

---

Generated with xum • Model: anthropic:claude-opus-5-5 • Thinking: max • Cost: $7.24