feat(routing): high tier escalates over task rules; benchmark generates hard prompts that stay hard

Routing semantics fix observed live: a task rule ("code → cheap model") was
pinning even high-complexity prompts to the cheap model — the tier system
never got a say. At tier=high the tier chain now leads and the task rule
demotes to fallback; low/medium keep task-rule-first. Unit-tested both ways.

The benchmark's hard probes are now GENERATED per run from two pools
(science: protein folding / telescope scheduling / materials discovery;
ops: epidemic logistics / grid markets / climate risk) with rotating domains
and numbers. Two findings against the live router drove this: ambiguity
carries the biggest weight in the complexity score, so interdisciplinary
prompts rate high reliably (0.52-0.66) while single-domain "hard" prompts
classify too confidently and sink to medium; and any fixed wording loses
novelty on every repeat. Their labels are null — straddling task labels is
precisely what makes them high, so classification is not judged.

Verified end-to-end: full run 100% answers, 13/13 classification, both hard
probes high → routed to the expensive model (one of them over a code task
rule), at $0.0018 each thanks to the tiny route-judge token budget.

web/js/views/benchmark.js

@@ -35,20 +35,36 @@ const SUITE = [
   { id: "chitchat", cat: "chitchat", label: ["chitchat"], quick: true, max: 48, prompt: "Hi! How are you today?", check: [] },
   // hard probes: dense enough that the router's complexity model rates them
   // high tier — this is where tier-based routing actually earns its keep.
-  // Each run picks a random variant: the router treats familiar prompts as
-  // less novel (by design), so a fixed wording would decay to medium/low.
-  { id: "hard-compiler", cat: "hard", label: ["code", "qa"], judge: "route", max: 8, check: [], variants: [
-    "Design an SSA-based optimizing compiler backend for an exposed-pipeline VLIW media processor: formulate instruction scheduling as an integer linear program with resource and hazard constraints, apply modulo scheduling to software-pipelined inner loops using rotating register files, do spill-aware register allocation via graph coloring with live-range splitting, add a translation-validated peephole pass, and explain how to prove the scheduler can never violate a structural hazard.",
-    "Build the code-generation backend of a JIT for GPU warp-specialized kernels: cost-model-driven instruction selection across divergent branches, software pipelining of memory latency with double-buffered shared memory, register allocation under occupancy pressure with rematerialization, a verified peephole pass over PTX, and a proof sketch that no schedule can deadlock the warp scheduler.",
-    "Construct an optimizing backend for a RISC-V vector coprocessor with configurable lane widths: vector-length-agnostic instruction selection as cost-model search, an integer-linear-program schedule around a chained load-store unit's structural hazards, VL-dependent spill-cost register allocation with live-range splitting, and a translation-validated peephole pass, plus the argument that the scheduler never issues into an occupied pipeline slot.",
-  ] },
-  { id: "hard-protocol", cat: "hard", label: ["qa", "code"], judge: "route", max: 8, check: [], variants: [
-    "Design and security-review an end-to-end encrypted coordination protocol for a federated fleet of 100000 IoT devices: lightweight key agreement under 32KB RAM budgets, per-group sender-key trees with post-compromise security, secure firmware-update attestation chained to a transparency log, metadata minimization against a network-level adversary, and revocation of physically captured devices without a round-trip to every peer. Analyze the trade-offs precisely, then specify the exact message flow for rotating a compromised region's keys.",
-    "Design an end-to-end encrypted messaging system for a disaster-relief mesh network that partitions for days: store-and-forward double ratchet with bounded state, group membership changes that commute across partitions, post-compromise security once connectivity returns, spam and sybil resistance without a central server, and forensic-resistant local storage. Weigh the trade-offs, then give the exact key schedule for merging two long-partitioned group histories.",
-    "Architect the security layer of a multi-tenant confidential-computing platform: remote attestation chained through TEE quotes to a transparency log, per-tenant key hierarchies with crypto-agile rotation, sealed storage that survives enclave upgrades, side-channel hardening policy for co-resident tenants, and revocation that cannot be rolled back by a malicious host. Analyze each trade-off, then specify the exact handshake for migrating a sealed workload between hosts.",
-  ] },
+  // Prompts are GENERATED fresh each run (domain + numbers rotate): the
+  // router deliberately scores familiar text as less novel, so any fixed
+  // wording would decay to medium/low over repeated runs.
+  // label: null — ambiguity between task labels is exactly what makes these high
+  { id: "hard-science", cat: "hard", label: null, judge: "route", max: 8, check: [], gen: "science" },
+  { id: "hard-ops", cat: "hard", label: null, judge: "route", max: 8, check: [], gen: "ops" },
 ];
 
+// Fresh hard prompts per run. Two findings drive the design (measured against
+// the live router): ambiguity carries the biggest weight in the complexity
+// score, so INTERDISCIPLINARY prompts (science + engineering + planning) rate
+// high reliably, while single-domain ones classify too confidently and drop
+// to medium; and familiar text loses novelty on every repeat, so domains and
+// numbers rotate to keep each run lexically fresh.
+function hardPrompt(kind) {
+  const pick = (a) => a[(Math.random() * a.length) | 0];
+  const n = (lo, hi) => lo + ((Math.random() * (hi - lo)) | 0);
+  const science = [
+    () => `Architect a protein-folding inference pipeline for a wet lab: ensemble MSA generation with taxonomy-aware subsampling, template selection under a ${n(30, 60)} percent identity ceiling, distributed attention across ${n(4, 16)} GPUs with activation checkpointing, confidence calibration against experimental B-factors, and an active-learning loop that decides which of ${n(300, 2000)} candidate mutants to synthesize next week.`,
+    () => `Schedule observations for a network of ${n(8, 30)} robotic telescopes chasing gravitational-wave counterparts: probability skymaps updated every ${n(5, 20)} minutes, slew-time-aware target assignment with decaying rewards, cloud-cover forecasts folded in as stochastic availability, spectroscopy triggered only above a ${n(40, 85)} percent kilonova likelihood, and a plan to validate the scheduler against the one historical multi-messenger event.`,
+    () => `Run a closed-loop materials-discovery campaign for solid-state electrolytes: DFT screening of ${n(2000, 9000)} candidates, Bayesian optimization driving ${n(3, 12)} parallel synthesis robots, CNN-based XRD phase identification, an ionic-conductivity target of ${n(2, 25)} mS/cm, and an explicit decision policy for when simulation is trusted over synthesis.`,
+  ];
+  const ops = [
+    () => `Model a dengue outbreak across ${n(20, 90)} districts by coupling an SEIR compartment model to vaccine cold-chain optimization: transmission parameters estimated from noisy case counts with a particle filter, uncertainty propagated into weekly demand forecasts, refrigerated routing under a ${n(120, 400)}-dose-per-crate constraint, and a defended ethical weighting for which ${n(3, 12)} districts get priority when supply covers ${n(20, 60)} percent of demand.`,
+    () => `Co-optimize a regional power grid's day-ahead market with a fleet of ${n(200, 2000)} grid-scale batteries: bid curves learned from ${n(2, 6)} years of price data, unit commitment as mixed-integer programming with ramp constraints, degradation amortized into marginal costs, N-1 contingency screening after every dispatch, and a mechanism-design argument that the bidding rules cannot be gamed by a coalition controlling ${n(10, 30)} percent of storage.`,
+    () => `Couple regional climate downscaling at ${n(2, 12)} km resolution to catastrophe-bond pricing: bias-correct ${n(10, 40)} CMIP ensemble members, generate stochastic hurricane tracks, propagate landfall distributions into a portfolio of ${n(30, 200)} bonds, price the attachment probabilities, and defend the discount rate applied to century-scale tail risk.`,
+  ];
+  return pick(kind === "science" ? science : ops)();
+}
+
 const LS_CUSTOM = "allaigate_bench_custom";
 const LS_HISTORY = "allaigate_bench_history";
 
@@ -168,7 +184,7 @@ export async function renderBenchmark() {
       for (let i = 0; i < tests.length; i++) {
         if (stopFlag) break;
         const test = tests[i];
-        const prompt = test.variants ? test.variants[(Math.random() * test.variants.length) | 0] : test.prompt;
+        const prompt = test.gen ? hardPrompt(test.gen) : test.prompt;
         mount(
           progressHost,
           h("div", { class: "card" },