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

[VickyXAI]

Router Core was two commits behind upstream (d7bc10c → 5ee7c23), plus a separate bug in the SDK-only free profile.

Off-catalog routing

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 leave every rung, fallbacks included.

Measured against the live catalog, before and after:

profile	off-catalog rungs per decision (before)	after
auto	3	0
eco	3	0
premium	3	0
free	3	0

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. Primaries moved only where portfolio.py already holds calibration evidence for the successor. Promotion waits for a calibration run; version recency is not a quality signal.

Priors come with it: 66 model profiles (was 30) and a 71-model capability snapshot. Two stale capability values had been reaching the hard filter — Haiku 4.5 at 8K max output (actually 64K), Sonnet 4.6 at 200K context (actually 1M).

config.py was not hand-transcribed. A TS→Python converter was validated by running it over config.ts at the old pin and confirming it reproduced the current config.py tier chains line for line (223/223) before being pointed at HEAD.

routing_profile="free" had collapsed to one model

Four of the five ids in FREE_TIERS were retired by NVIDIA, and they were every primary plus all but one fallback. Every tier fell through to a single survivor with zero fallbacks.

Nothing looked broken because the gateway server-redirects a retired free id — the 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:

nvidia/step-3.7-flash          -> nvidia/nemotron-3.5-lightning
nvidia/nemotron-nano-9b-v2     -> nvidia/nemotron-3-nano-30b
nvidia/nemotron-nano-12b-v2-vl -> nvidia/nemotron-3-nano-30b
nvidia/mistral-nemotron        -> nvidia/nemotron-3-nano-30b

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. Worth reporting upstream: 5ee7c23 kept nano-omni as the one survivor of its own 2026-08-30 pass.

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.

The new depth test asserts fallback depth per tier rather than membership. The table stayed internally consistent all through the rot, so a membership check could never have caught it.

Verification

• 669 unit tests pass, including the 88 cross-language parity decisions against the upstream frozen corpus
• black --check and ruff check clean
• mypy reports the same 198 baseline errors before and after — no regression
• End-to-end with no wallet: smart_chat(routing_profile="free") returns nemotron-3-nano-30b and nemotron-3.5-lightning, $0