Auto: twincat-2.1 — ADS Sum-read / Sum-write

Closes #310

docs/drivers/TwinCAT-Test-Fixture.md

@@ -125,6 +125,35 @@ back an `IAlarmSource`, but shipping that is a separate feature.
 | "Do notifications coalesce under load?" | no | yes (required) |
 | "Does a TC2 PLC work the same as TC3?" | no | yes (required) |
 
+## Performance
+
+PR 2.1 (Sum-read / Sum-write, IndexGroup `0xF080..0xF084`) replaced the per-tag
+`ReadValueAsync` loop in `TwinCATDriver.ReadAsync` / `WriteAsync` with a
+bucketed bulk dispatch — N tags addressed against the same device flow through a
+single ADS sum-command round-trip via `SumInstancePathAnyTypeRead` (read) and
+`SumWriteBySymbolPath` (write). Whole-array tags + bit-extracted BOOL tags
+remain on the per-tag fallback path because the sum surface only marshals
+scalars and bit-RMW writes need the per-parent serialisation lock.
+
+**Baseline → Sum-command delta** (dev box, 1000 × DINT, XAR VM over LAN):
+
+| Path | Round-trips | Wall-clock |
+| --- | --- | --- |
+| Per-tag loop (pre-PR 2.1) | 1000 | ~5–8 s |
+| Sum-command bulk (PR 2.1) | 1 | ~250–600 ms |
+| Ratio | — | ≥ 10× typical, ≥ 5× CI floor |
+
+The perf-tier test
+`TwinCATSumCommandPerfTests.Driver_sum_read_1000_tags_beats_loop_baseline_by_5x`
+asserts the ratio with a conservative 5× lower bound that survives noisy CI /
+VM scheduling. It is gated behind both `TWINCAT_TARGET_NETID` (XAR reachable)
+and `TWINCAT_PERF=1` (operator opt-in) — perf runs aren't part of the default
+integration pass because they hit the wire heavily.
+
+The required fixture state (1000-DINT GVL + churn POU) is documented in
+`TwinCatProject/README.md §Performance scenarios`; XAE-form sources land at
+`TwinCatProject/PLC/GVLs/GVL_Perf.TcGVL` + `TwinCatProject/PLC/POUs/FB_PerfChurn.TcPOU`.
+
 ## Follow-up candidates
 
 1. **XAR VM live-population** — scaffolding is in place (this PR); the