feat: surface per-model cost and limit from static catalog

[manojp99]

Summary

Re-enables cost and limit fields in build_provider_block(). Both were stripped earlier to dodge opencode's silent config-decode failure mode (a single malformed entry wipes the whole provider's models block).

Why a static catalog is necessary

opencode bundles the models.dev registry — the catalog has rich pricing for canonical provider IDs (anthropic, openai, gemini, etc.). But opencode keys catalog lookups by (providerID, modelID), not just modelID. Our custom amplifier provider serves the same model IDs Anthropic uses (claude-haiku-4-5, claude-sonnet-4-6, ...) but opencode treats them as un-cataloged because the providerID doesn't match.

Empirically confirmed: Without per-model cost in our config, opencode renders every turn at $0.00:

// our config (limit only):
"claude-haiku-4-5-20251001": {"name": "...", "limit": {...}}

// opencode's loaded state:
"cost": {"input": 0, "output": 0, "cache": {"read": 0, "write": 0}}    ← zeros

Two alternatives we tested first

Alternative	Result
opencode's provider.api: 'anthropic' field as catalog-inheritance hint	Did NOT cause inheritance. cost remained zeros after restart.
Submit amplifier to models.dev (sst/models)	External review cycle, and our served model set varies by amplifier-agent's host_config.json — can't declare a fixed list upstream.

Hardcoded catalog in the adapter is the self-contained path. Maintenance: when a provider changes prices, update MODEL_PRICING_PER_MILLION. Source of truth referenced inline: github.com/sst/models.

Generated config — before vs after

Before:

"claude-haiku-4-5-20251001": { "name": "Claude Haiku 4.5" }

After:

"claude-haiku-4-5-20251001": {
  "name":  "Claude Haiku 4.5",
  "cost":  { "input": 1.0, "output": 5.0, "cache_read": 0.1, "cache_write": 1.25 },
  "limit": { "context": 200000, "output": 64000 }
}

opencode normalizes flat cache_read/cache_write to nested cache: {read, write} internally — confirmed via /provider endpoint probe.

Initial coverage

Provider	Models	Source
Anthropic	Claude Haiku 4.5, Sonnet 4.6, Opus 4.8	models.dev / Anthropic pricing page
OpenAI	Placeholder (commented) — fill when amplifier-agent serves them	n/a

Compatibility with the cost telemetry path

microsoft/amplifier-agent#68 surfaces real per-turn cost_usd on the chat-completions response (Decimal-as-string, from provider modules' compute_cost). That ships on the wire today; this PR is for opencode's TUI picker display (per-model rate) and the rolling cost calc opencode does from tokens × rates. The two complement each other:

• wire cost_usd: authoritative $$ for cost-aware clients (curl, custom wrappers, telemetry tools)
• catalog rates (this PR): opencode's TUI rendering