fix(routing): re-sync Router Core to 5ee7c23, and rebuild the free tier from probed ids

CHANGELOG.md

@@ -2,6 +2,60 @@
 
 All notable changes to blockrun-llm will be documented in this file.
 
+## Unreleased
+
+### Changed
+- **Router Core re-synced to upstream `5ee7c23`** (was `d7bc10c`, two commits
+  behind — the same pin the TypeScript SDK bundles). Upstream V3.5 rebuilds
+  every tier chain around ids the public catalog actually lists: the withheld
+  `kimi-k2.5/k2.6/k2.7`, both grok-4-fast pairs, `grok-4-0709`,
+  `claude-opus-4.6` and `gemini-3-pro-preview` are gone from every rung,
+  including fallbacks, so a routed model is always one a user can find on
+  blockrun.ai/models. Before this sync every profile carried 3–4 off-catalog
+  rungs per decision; now it carries none.
+
+  Primaries moved only where `portfolio.py` already holds calibration evidence
+  for the successor — Gemini 3.5 Flash where Kimi K2.7 was, GPT-5 Mini for
+  agentic MEDIUM, Sonnet 5 over Sonnet 4.6, DeepSeek Reasoner for the cheap
+  reasoning head. The newer generation the catalog already sells enters as
+  fallback rungs: GPT-5.6 Luna/Terra, Gemini 3.6 Flash and 3.5 Flash-Lite,
+  GLM-5.3 and 5.3-Flash, Grok 4.3 and 4.5, Kimi K3, Qwen 3.7 Plus, MiniMax M3.
+  Promotion waits for a calibration run, because version recency is not a
+  quality signal. The expired GLM-5.1 promotion is dropped; the promotions
+  mechanism stays wired with an empty list.
+
+- **Routing priors regenerated: 66 model profiles** (was 30) and a **71-model
+  capability snapshot** (`model_capabilities.py`), both from upstream's probe
+  and catalog-sync scripts. Two stale capability values had been reaching the
+  hard filter: Haiku 4.5 at 8K max output (actually 64K) and Sonnet 4.6 at 200K
+  context (actually 1M). Kimi K3 replaces K2.7 in the Mandarin extraction band,
+  widened to 0.12 so the auto affinity floor gap (0.10) cannot let price
+  re-select a non-native model, and K3 joins the extraction evidence pool since
+  it is no longer on the MEDIUM chain.
+
+### Fixed
+- **`routing_profile="free"` had collapsed to a single model with no
+  fallbacks.** Four of the five ids in the SDK-only `FREE_TIERS` table —
+  `nvidia/step-3.7-flash`, `nemotron-nano-9b-v2`, `mistral-nemotron`,
+  `nemotron-nano-12b-v2-vl` — were retired by NVIDIA, and they were every
+  primary plus all but one fallback. Nothing looked broken because the gateway
+  server-redirects a retired free id: a dead rung returns 200 and answers
+  normally, while quietly serving a different model. That is the shape that
+  defeats a host's `unavailable_models`, and it is why the table rotted
+  unnoticed.
+
+  The table is rebuilt from ids verified by a two-pass probe that reads back
+  the response's own `model` field, since a 200 proves nothing. Two models the
+  catalog still prices at $0 did not survive that check and are excluded:
+  `nemotron-3-ultra-550b` and `nemotron-3-nano-omni-30b-a3b-reasoning` both
+  answer as `nemotron-3-nano-30b`. Every tier is back to four candidates, and
+  the free tier is no longer NVIDIA-only — `cohere/north-mini-code` and
+  `poolside/laguna-xs-2.1` serve at $0 and carry the free coding load.
+
+  A new test asserts fallback *depth* per tier, not just membership: the table
+  stayed internally consistent all through the rot, so membership alone could
+  never have caught it.
+
 ## 1.13.0 — 2026-08-26
 
 ### Fixed

blockrun_llm/router_adapter.py

@@ -61,47 +61,59 @@ class ResolvedRoutingDecision(RoutingDecision, total=False):
 #: The BlockRun free tier is a gateway concept, not a Router Core profile: the
 #: core's tiers rank paid models by task affinity, and its evidence candidates
 #: are paid ids. ``routing_profile="free"`` therefore routes on the rules
-#: strategy over this NVIDIA-only tier table, and the adapter additionally
-#: drops any candidate the catalog does not price at $0.
+#: strategy over this tier table, and the adapter additionally drops any
+#: candidate the catalog does not price at $0.
 #:
-#: Refreshed 2026-08-15 against the live catalog. NVIDIA has EOL'd (HTTP 410)
-#: the free DeepSeek family — ``deepseek-v4-flash`` was the last to go on
-#: 2026-08-12 — plus ``llama-4-maverick`` and the qwen3 SKUs, which is what the
-#: previous table pointed at. ``gpt-oss-120b/20b`` stay out of the primaries:
-#: they are hidden from ``/v1/models`` (so they carry no catalog price) over
+#: Refreshed 2026-08-31. Every id below was verified by asking the gateway for
+#: it twice and reading back the ``model`` field of the reply, because a 200 is
+#: not proof: blockrun server-redirects a retired free id to a live one, so a
+#: dead rung answers normally while quietly serving something else. That is the
+#: shape that defeats a host's ``/exclude``, and it is why the previous table
+#: went stale unnoticed. Substituting on 2026-08-31, hence absent here:
+#: ``step-3.7-flash``, ``nemotron-nano-9b-v2``, ``nemotron-nano-12b-v2-vl`` and
+#: ``mistral-nemotron`` (retired upstream 2026-08-30), plus ``nemotron-3-ultra-550b``
+#: and ``nemotron-3-nano-omni-30b-a3b-reasoning`` — the latter two still list at
+#: $0 in ``/v1/models`` but both answer as ``nemotron-3-nano-30b``.
+#:
+#: The table is no longer NVIDIA-only: ``cohere/north-mini-code`` and
+#: ``poolside/laguna-xs-2.1`` serve at $0 and carry the free coding load.
+#: ``gpt-oss-120b/20b`` stay out — proxy-only ids with no catalog price, under
 #: the NVIDIA free tier's prompt-retention policy.
 FREE_TIERS: dict[str, TierConfig] = {
     "SIMPLE": {
-        "primary": "nvidia/step-3.7-flash",  # 131K ctx, fast general chat + reasoning
+        # Fastest free model (~121 tok/s), and latency is the only axis that
+        # separates free rungs — they all cost $0.
+        "primary": "nvidia/nemotron-3-nano-30b",  # 131K ctx
         "fallback": [
-            "nvidia/nemotron-nano-9b-v2",
-            "nvidia/mistral-nemotron",
-            "nvidia/nemotron-3-nano-omni-30b-a3b-reasoning",
+            "nvidia/nemotron-3.5-lightning",
+            "nvidia/llama-3.2-11b-vision",
+            "poolside/laguna-xs-2.1",
         ],
     },
     "MEDIUM": {
-        "primary": "nvidia/step-3.7-flash",
+        "primary": "nvidia/nemotron-3.5-lightning",  # 1M ctx — free tier flagship
         "fallback": [
-            "nvidia/mistral-nemotron",
-            "nvidia/nemotron-3-nano-omni-30b-a3b-reasoning",
-            "nvidia/nemotron-nano-9b-v2",
+            "nvidia/nemotron-3-nano-30b",
+            "poolside/laguna-xs-2.1",
+            "cohere/north-mini-code",
         ],
     },
     "COMPLEX": {
-        # Largest free context (256K) and the only free vision model, so it also
-        # absorbs long or multi-modal requests.
-        "primary": "nvidia/nemotron-3-nano-omni-30b-a3b-reasoning",
+        # Only free model above 256K, so it absorbs long inputs; the vision rung
+        # behind it absorbs multi-modal ones.
+        "primary": "nvidia/nemotron-3.5-lightning",
         "fallback": [
-            "nvidia/step-3.7-flash",
-            "nvidia/mistral-nemotron",
-            "nvidia/nemotron-nano-12b-v2-vl",
+            "cohere/north-mini-code",  # 256K ctx
+            "nvidia/nemotron-3-nano-30b",
+            "nvidia/llama-3.2-11b-vision",  # free vision
         ],
     },
     "REASONING": {
-        "primary": "nvidia/nemotron-3-nano-omni-30b-a3b-reasoning",
+        "primary": "nvidia/nemotron-3.5-lightning",
         "fallback": [
-            "nvidia/step-3.7-flash",
-            "nvidia/nemotron-nano-9b-v2",
+            "nvidia/nemotron-3-nano-30b",
+            "cohere/north-mini-code",
+            "poolside/laguna-xs-2.1",
         ],
     },
 }

blockrun_llm/router_core/__init__.py

@@ -2,7 +2,7 @@
 Router Core — deterministic, constraint-first model routing.
 
 Python port of `@blockrun/router-core <https://github.com/BlockRunAI/router-core>`_
-(upstream commit ``d7bc10c``), the same routing engine the TypeScript SDK and
+(upstream commit ``5ee7c23``), the same routing engine the TypeScript SDK and
 the BlockRun gateway use. The package is deliberately product-neutral: task
 classification, hard capability filtering, portfolio scoring, ordered
 fallbacks, and routing configuration. It contains no wallet, gateway client,