fix(bench-intents): bracket more than 26 candidate intents instead of failing

CHANGELOG.md

@@ -7,6 +7,11 @@ changes, new features. **Patch:** fixes. Images are published to
 a release is validated. Release process: [runbook](docs/operate/runbook.md#release-a-new-version). Gateway/CLI-only releases
 (`make release-gateway`) reuse the previous serving images.
 
+## Unreleased
+
+- `bench-intents`: more than 26 candidate intents run as a 2-stage bracket (groups of ≤ 20, top 5 to a final round)
+  instead of being rejected by the server (`--dataset banking77` sends 30).
+
 ## v0.1.2 (2026-09-28)
 
 Gateway and monitoring; serving images unchanged (v0.1.0).

benchmarks/runs/20260928-v010-verification/README.md

@@ -12,7 +12,7 @@ committed receipts in `benchmarks/` (earlier images and hardware) and with the i
 | `bench-calibration` | 50 | 46 (92%) | 43–46 (image parity) | Pass |
 | `bench` (triage, code review, security) | 30 | 22 (73.3%) | 23–24 on L4 receipts | Pass (±2 items is noise on 30) |
 | `bench-intents` clinc150 (small set) | 30 | 29 (96.7%) after tie-break; domain 28/30 | Committed receipt 96.7% | Pass |
-| `bench-intents` banking77 (small set) | 30 | 0: every request rejected, `at most 26 alternatives` | Committed receipt (Sep 20, older server) | **Harness issue**, see below |
+| `bench-intents` banking77 (small set) | 30 | 0 at first (every request rejected, `at most 26 alternatives`); after the bracket fix: **23 (76.7%)** | Committed receipt 26 (Sep 20, older server allowed 30 options in one question) | Fixed (harness) |
 | `bench-bbox` | 12 | mIoU 0.406 argmax / 0.422 expectation | 0.290 / 0.377 (Cloud Run L4) | Pass |
 | `bench-rerank` | 30 queries | nDCG@10 0.824, 0.856, 0.856 (3 runs); 0% ties, 100% poison quarantine | 0.9265 Cloud Run L4, 0.8502 Vertex L4 | Pass (in the range of earlier receipts) |
 | `bench-permutation` | 16 | 100% accuracy, 6.25% flip rate | EXP-13 | Pass |
@@ -21,9 +21,11 @@ committed receipts in `benchmarks/` (earlier images and hardware) and with the i
 | `bench-ecotone` | — | Not run (needs the C++ ecotone sidecar) | — | Skipped |
 
 **Two findings, both pre-existing, not regressions:**
-- `bench-intents --dataset banking77` sends all 30 candidate intents in one `choice` question. The server has
+- `bench-intents --dataset banking77` sent all 30 candidate intents in one `choice` question. The server has
   rejected more than 26 options since before v0.1.0 (`ab208dd` had the same check); the committed receipt predates
-  it. The harness should split into brackets (as `dgem systemone serve` does) or use the 26-option slice.
+  it. **Fixed:** with more than 26 options the harness now runs a 2-stage bracket (groups of ≤ 20, top 5 of each
+  group to a final round), like `dgem systemone serve`. Result 23/30 (76.7%); keeping the top 3 per group gave 22,
+  the top 8 gave 23. The misses are confusable intent pairs (e.g. `card_arrival` vs `card_delivery_estimate`).
 - `intent_banking77`, `intent_clinc150`, `intent_tiebreak`, `jevbench_generic` and `tn_disambiguation` are harness
   templates: they take option lists from `bench-intents`, `bench-jev` or `bench-ecotone`, so the gateway's sample
   variables leave them with fewer than two options (400). The harnesses themselves pass (above).