refactor(router): unified model catalog as single source of truth

[steventohme]

Summary

Introduces internal/router/catalog/ — one struct literal per model holding tier, ordered provider bindings, and per-binding pricing. pricing and capability become thin facades over catalog, preserving every existing public type and helper.

Motivation

Adding a single new model used to require touching up to 11 places:

• internal/router/pricing/pricing.go table
• internal/router/capability/tier.go tiers + tier_test.go
• internal/router/cluster/artifacts/v0.X/model_registry.json
• internal/router/cluster/artifacts/v0.X/rankings.json
• internal/router/cluster/artifacts/v0.X/metadata.yaml
• install/install.sh, install/cc-statusline.sh
• the per-provider modelIDMap in cmd/router/main.go (when SOC 2 routing flips a model)

It's all the same data structurally — exactly the smell aider, litellm, Cline, and continue.dev solve with a single typed catalog. This PR collapses the Go-side pieces of that pile to one struct literal in catalog.go plus a go run ./cmd/genprices to regenerate the shell price block.

Multi-provider per model (PR 8.5 of the SOC 2 plan)

Each Model carries an ordered Providers []ProviderBinding list. The cluster scorer picks the first binding whose Provider name is in the deploy's available set, and the planner can ask the catalog for (provider, model)-keyed pricing. Today every entry is a single-element list — the shape is in place so the SOC 2 isolation work (#187) can append an OpenRouter fallback to each OSS row without touching any call site:

{ID: "deepseek/deepseek-v4-pro", Tier: TierHigh, Providers: []ProviderBinding{
    {Provider: ProviderFireworks,  UpstreamID: "...", Price: Pricing{...}},
    {Provider: ProviderOpenRouter,                    Price: Pricing{...}}, // self-hoster fallback
}},

Managed prod (no OPENROUTER_API_KEY) gets the primary binding; self-hosters with only an OpenRouter key get the trailing one.

What changed

• New package internal/router/catalog/ with catalog.go (data) + lookup.go (accessors) + catalog_test.go (schema invariants).
• internal/router/pricing/pricing.go — table deleted; For/All/Pricing/DefaultCacheReadMultiplier now thin pass-throughs into catalog. Math helpers (EffectiveInputCost, EffectiveOutputCost) unchanged.
• internal/router/capability/tier.go — tiers map deleted; Tier/TierLow/TierMid/TierHigh aliased from catalog; TierFor/IsAtOrBelow/AllowedAtOrBelow/Validate delegate.
• cmd/genprices already reads through otel.AllPricing() → pricing.All() → catalog. Zero-diff regen.
• CLAUDE.md + AGENTS.md mirrors updated for internal/router/, pricing/, capability/, and the new catalog/.

Behavior- and price-preserving

• pricing.For / pricing.All / pricing.EffectiveInputCost / EffectiveOutputCost API-stable; values byte-identical to pre-refactor.
• capability.TierFor / IsAtOrBelow / AllowedAtOrBelow / Validate API-stable.
• install/install.sh + install/cc-statusline.sh regenerate with zero diff.

What's NOT in this PR

• The planner call-site switch to catalog.PriceFor(provider, model) and the cluster scorer's ResolveBinding plumbing — both land in the follow-up PR that actually introduces multi-binding rows (SOC 2 provider isolation).
• internal/router/model.go ModelSpec (wire-format capabilities like CapReasoning, CapExtendedThinking) — separate concept, not folded here.
• No retraining, no artifact version bump.

Test plan

• go test ./... green
• go run ./cmd/genprices produces zero diff vs current install/ files
• New catalog_test.go invariants: no dup IDs, every model has ≥1 binding, every Provider is a canonical constant, every binding has positive pricing, ByID date-suffix fallback works
• Staging boot: cluster scorer initializes against catalog-backed pricing/tier; no Validate() regressions

🤖 Generated with Claude Code

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit a636492. Configure here.}

[cursor]

OTel span attribute keys silently renamed breaking downstream consumers

High Severity

The OTel span attribute keys were renamed from pricing.requested_input_per_1m / pricing.actual_input_per_1m etc. to catalog.requested_input_per_1m / catalog.actual_input_per_1m. These are wire-protocol names consumed by external dashboards, alerting rules, and analytics queries. Any downstream consumer querying the old pricing.* keys will silently receive no data. The PR claims to be "behavior- and price-preserving" but these attribute key names are part of the observable contract. The attribute keys are an external interface, not an internal Go package name — they needn't track the refactor.

Additional Locations (2)

• internal/proxy/gemini.go#L113-L117
• internal/proxy/service.go#L1506-L1510

 

^{Reviewed by Cursor Bugbot for commit a636492. Configure here.}