fix(chat): reserve $0 for the hidden NVIDIA ids the account sheet bills at $0

src/utils/constants.ts

@@ -256,7 +256,7 @@ export type RoutingMode = keyof typeof MODEL_TIERS;
  * promise "still free when it returns".
  *
  * What that check cannot cover, and does not pretend to: a member the catalogue
- * does not list at all. Eight of these are in that state on both gateways today,
+ * does not list at all. Twenty-three of these are in that state on both gateways today,
  * and delisting is not death — gpt-oss-120b is the gateway's own free fallback
  * and answers for itself while absent from GET /v1/models. The sweep names each
  * one as UNVERIFIED rather than counting it as checked; the only way to settle
@@ -288,6 +288,32 @@ export const FREE_CHAT_MODELS: ReadonlySet<string> = new Set<string>([
   "nvidia/nemotron-3.5-lightning", // itself on sol only; super-120b on Base
   "nvidia/muse-glimmer-30b", // now listed on both; itself on sol only
   "nvidia/gemma-4-31b", // delisted on both; redirected to super-120b (gateway MODEL_REDIRECTS)
+  // Retired/hidden NVIDIA ids the ACCOUNT rail still prices at $0
+  // (blockrun.ai/api/pricing, billingMode "free", 2026-10-20) and the wallet
+  // gateways redirect onto a $0 model (MODEL_REDIRECTS / fallbackModel). All
+  // fifteen kept here answered 200 with no payment header on BOTH chains on
+  // 2026-10-20, served by nemotron-3-super-120b, gpt-oss-20b, nano-omni or
+  // llama-3.2-11b-vision. Not routed — an explicit call just reserves $0
+  // instead of the $5/$30 default (verify:prices flagged each one).
+  "nvidia/nemotron-3-super-120b-a12b-free",
+  "nvidia/qwen3.5-397b-a17b",
+  "nvidia/nemotron-ultra-253b",
+  "nvidia/nemotron-super-49b",
+  "nvidia/deepseek-v3.2",
+  // NOT nvidia/deepseek-v4-pro or nvidia/deepseek-v4-flash: both are being
+  // removed from the gateways (blockrun#848 for -flash, live on blockrun.ai;
+  // -pro approved 2026-10-20) and then answer an unknown-model 400 — a $0
+  // reserve for an id that no longer exists would only hide that.
+  "nvidia/nemotron-3-nano-omni-nim",
+  "nvidia/mistral-large-3-675b",
+  "nvidia/qwen3-coder-480b",
+  "nvidia/devstral-2-123b",
+  "nvidia/glm-4.7",
+  "nvidia/llama-4-maverick",
+  "nvidia/qwen3-next-80b-a3b-instruct",
+  "nvidia/qwen3.5-122b-a10b",
+  "nvidia/seed-oss-36b",
+  "nvidia/mistral-small-4-119b",
 ]);
 
 /**
@@ -589,19 +615,30 @@ export const GATEWAY_CHARS_PER_TOKEN_OBSERVED = 2.08;
  * chat-anthropic.ts already normalised bare -> prefixed for the LEDGER; this is
  * the same normalisation for the GATE, which is where the money decision is.
  * Post-slash segments are unique across the catalog (pinned by a test), so the
- * mapping is unambiguous.
+ * mapping is unambiguous — with one deliberate exception class: a $0 NVIDIA
+ * mirror whose bare name is also a PAID model's (nvidia/deepseek-v4-pro vs
+ * deepseek/deepseek-v4-pro was the case, until the mirror was retired on
+ * 2026-10-20; none remain). There the paid id must own the bare spelling:
+ * mapping it to the free mirror would reserve $0 for a paid call, the gate
+ * bypass this file exists to prevent. Free ids go in FIRST so any paid id
+ * with the same bare name overwrites them.
  */
 const BARE_TO_PREFIXED: Record<string, string> = Object.fromEntries(
   [
+    ...FREE_CHAT_MODELS,
     ...Object.keys(CHAT_PRICE_PER_MTOKEN),
     ...Object.keys(LONG_CONTEXT_PRICE_PER_MTOKEN),
     ...Object.values(MODEL_TIERS).flat(),
-    ...FREE_CHAT_MODELS,
   ]
     .filter((id: string) => id.includes("/"))
     .map((id: string) => [id.slice(id.indexOf("/") + 1), id]),
 );
 
+/** $0 ids whose bare spelling belongs to a paid model (see BARE_TO_PREFIXED). */
+export const FREE_SHADOWED_BY_PAID: ReadonlySet<string> = new Set(
+  [...FREE_CHAT_MODELS].filter((id) => BARE_TO_PREFIXED[id.slice(id.indexOf("/") + 1)] !== id),
+);
+
 export function canonicalChatModel(model: string): string {
   const id = model.trim();
   if (Object.hasOwn(CHAT_PRICE_PER_MTOKEN, id)) return id;

test/chat.test.ts

@@ -3,7 +3,7 @@ import { test } from "node:test";
 import assert from "node:assert/strict";
 import { estimateChatCost, freeTierTruncationNote } from "../src/tools/chat.js";
 import { handleAnthropicNative, anthropicCallCost } from "../src/tools/chat-anthropic.js";
-import { MODEL_TIERS, CHAT_PRICE_PER_MTOKEN, DEFAULT_CHAT_PRICE, FREE_CHAT_MODELS } from "../src/utils/constants.js";
+import { MODEL_TIERS, CHAT_PRICE_PER_MTOKEN, DEFAULT_CHAT_PRICE, FREE_CHAT_MODELS, FREE_SHADOWED_BY_PAID } from "../src/utils/constants.js";
 import { withTxFee } from "../src/utils/tx-fee.js";
 import type { BudgetState } from "../src/types.js";
 
@@ -203,8 +203,14 @@ test("every MODEL_TIERS.free entry is in FREE_CHAT_MODELS, and every member rese
   }
   for (const m of FREE_CHAT_MODELS) {
     assert.equal(estimateChatCost(1024, undefined, m, undefined, 600 * 1024), 0, m);
-    // The bare spelling is a real, chargeable-or-free id too (see BARE_TO_PREFIXED).
-    assert.equal(estimateChatCost(1024, undefined, m.slice(m.indexOf("/") + 1), undefined, 600 * 1024), 0, `bare ${m}`);
+    // The bare spelling is a real, chargeable-or-free id too (see BARE_TO_PREFIXED)
+    // — unless a PAID model owns it, in which case it must reserve that model.
+    const bare = m.slice(m.indexOf("/") + 1);
+    if (FREE_SHADOWED_BY_PAID.has(m)) {
+      assert.ok(estimateChatCost(1024, undefined, bare, undefined, 600 * 1024) > 0, `bare ${m} belongs to a paid model`);
+    } else {
+      assert.equal(estimateChatCost(1024, undefined, bare, undefined, 600 * 1024), 0, `bare ${m}`);
+    }
   }
 });
 
@@ -366,8 +372,46 @@ test("every catalog id has a unique vendor-less segment — the mapping cannot b
     if (!byBare.has(bare)) byBare.set(bare, new Set());
     byBare.get(bare)!.add(id);
   }
-  const clashes = [...byBare.entries()].filter(([, set]) => set.size > 1);
+  // The one tolerated clash: a $0 NVIDIA mirror sharing a PAID model's name.
+  // The paid id owns the bare spelling (FREE_SHADOWED_BY_PAID), so the clash
+  // can only over-reserve, never reserve $0 for a paid call.
+  const clashes = [...byBare.entries()].filter(([, set]) => {
+    if (set.size < 2) return false;
+    const paid = [...set].filter((id) => !FREE_CHAT_MODELS.has(id));
+    const free = [...set].filter((id) => FREE_CHAT_MODELS.has(id));
+    return !(paid.length === 1 && free.every((id) => FREE_SHADOWED_BY_PAID.has(id)));
+  });
   assert.deepEqual(clashes, [], `two vendors ship the same model name: ${JSON.stringify(clashes)}`);
+  // nvidia/deepseek-v4-pro was the one case until it was retired (2026-10-20).
+  assert.deepEqual([...FREE_SHADOWED_BY_PAID], []);
+});
+
+// The account sheet (blockrun.ai/api/pricing) prices these hidden NVIDIA ids
+// at $0 and both wallet gateways answer them unpaid (2026-10-20); before this
+// an explicit call reserved the $5/$30 default and an exhausted budget refused it.
+test("account-sheet $0 NVIDIA ids reserve nothing, and are not routed", () => {
+  const ids = [
+    "nvidia/nemotron-3-super-120b-a12b-free", "nvidia/qwen3.5-397b-a17b", "nvidia/nemotron-ultra-253b",
+    "nvidia/nemotron-super-49b", "nvidia/deepseek-v3.2",
+    "nvidia/nemotron-3-nano-omni-nim", "nvidia/mistral-large-3-675b", "nvidia/qwen3-coder-480b",
+    "nvidia/devstral-2-123b", "nvidia/glm-4.7", "nvidia/llama-4-maverick", "nvidia/qwen3-next-80b-a3b-instruct",
+    "nvidia/qwen3.5-122b-a10b", "nvidia/seed-oss-36b", "nvidia/mistral-small-4-119b",
+  ];
+  for (const id of ids) {
+    assert.equal(estimateChatCost(1024, undefined, id, undefined, 100_000), 0, id);
+    assert.ok(!(MODEL_TIERS.free as readonly string[]).includes(id), `${id} must not join the routing order`);
+  }
+  // Retired from the gateways (unknown-model 400 since 2026-10-20): not free.
+  for (const gone of ["nvidia/deepseek-v4-pro", "nvidia/deepseek-v4-flash", "deepseek-v4-flash"]) {
+    assert.ok(!FREE_CHAT_MODELS.has(gone), gone);
+    assert.ok(estimateChatCost(1024, undefined, gone, undefined, 100_000) > 0, gone);
+  }
+  // The bare name of the paid DeepSeek still reserves the paid rate.
+  assert.equal(
+    estimateChatCost(1024, undefined, "deepseek-v4-pro", undefined, 100_000),
+    estimateChatCost(1024, undefined, "deepseek/deepseek-v4-pro", undefined, 100_000),
+  );
+  assert.ok(estimateChatCost(1024, undefined, "deepseek-v4-pro", undefined, 100_000) > 0);
 });
 
 // ── Two flagship ids landed above the default, and nothing noticed ──