BPMN eval: smart-triage + high-volume-batch flake on intrinsic weight after the hunt is removed (#3228)

[nikhil-maryala]

Context

smart-triage and high-volume-batch regressed to MAX_TURNS / turn_timeout in the eu nightly. The registry-hunt cause is fixed at the shared layer in #3228 (verified: the timed-out smart-triage run on that branch does 0 uip or / uip is connections / registry search commands, down from 16). But both tasks still flake over the 1800s turn budget with hunting fully removed — the agent deliberates through a large shape (extract/classify/two gateways/HITL/handlers/ends/diagram/package metadata for smart-triage; multi-instance + aggregate + escalation for high-volume-batch) and does not author in time on some runs.

Evidence: on #3228 (carve-out only), a smart-triage run timed out at 1803s with 0 hunt commands and 0 writes — 92 slow turns, never reached authoring. This is intrinsic task weight, not hunting.

Options to evaluate

1. Decompose the eval — split the heavy authoring into smaller graded steps, or provide a partial fixture so the agent authors a delta rather than the whole shape from scratch.
2. Right-size the budget — but test(maestro-bpmn): generous task/turn timeouts for 15 slow BPMN tasks #2940/test(uipath-maestro-bpmn): raise turn/max-turn budgets for 4 verified BPMN eval tasks #3150 already bumped these twice; a further raise chases eu per-turn latency (see [[coder-eval-task-timeout-config]]).
3. Tighten the authoring recipe — a point-of-use "author the whole shape in one pass from this skeleton" prompt/guide nudge (the fix(uipath-maestro-bpmn): scope metadata regeneration away from IS draft handoffs #3152 lesson: point-of-use changes behavior more than a referenced file).

Acceptance

smart-triage and high-volume-batch reach SUCCESS reliably (e.g. ≥4/5) on claude/sonnet-5, with the mechanism documented (decomposition vs recipe vs budget).

[nikhil-maryala]

Update: #3228 was extended to cover output-mapping enrichment (not just input discovery). That improved smart-triage from 0/2 → 2/3 on claude/sonnet-5 (SUCCESS 1112s/1704s, one timeout 1803s) — so the enrichment-flail portion of the weight is now handled at the shared layer. This issue is reduced to the residual: smart-triage (and high-volume-batch) still flake ~1/3 at the 1800s edge with both hunt and output-flail removed, purely on shape size. The remaining options are decomposition or a point-of-use one-pass authoring recipe (budget is off the table per prior direction).