feat(usage): user-configurable per-model cost overlay (providers.<name>.modelCosts) #1009

Why

The Logs ~$ estimate is priced from two compiled sources: the jawcode catalog
and the expected-price overlay. Both match exact native provider/model ids, so
internal and custom providers often fall through:

• provider/model ids from a proxy or internal gateway can be mangled or
  namespaced differently than the catalog rows (there is deliberately no
  fuzzy/case-fold matching), leaving ~$ as an em dash;
• even when an id matches, the operator's actual cost may vary from the catalog
  list price (reseller rates, negotiated pricing, pooled/subscription billing).

There is currently no way to set a price without editing source code.

Proposal

Add providers.<name>.modelCosts — a per-model display price overlay (USD per
1M tokens), following ocx's existing flat per-model config convention
(modelContextWindows, modelMaxInputTokens, ...):

"providers": {
  "blsc": {
    "adapter": "openai-chat",
    "baseUrl": "https://llmapi.blsc.cn",
    "modelCosts": {
      "deepseek-v4-flash": { "input": 0.14, "output": 0.28, "cacheRead": 0.0028, "cacheWrite": 0 }
    }
  }
}

Resolution order becomes: user-configured modelCosts → jawcode catalog →
expected-price overlay → model-level vendor fallback. All-zero entries fall
through ("zero = not billable here", consistent with catalog semantics).

Semantics

• Display-time estimation only; never billing.
• Rows are lifted from config at load and every persist path into a versioned
  registry; the estimator memo keys on that version, so edits apply without a
  restart and stale rows are never served.
• New provider_cost_overlay estimate reason surfaced in the Logs detail view
  (with i18n labels).
• Strict config validation (non-negative finite 4-tuples); the management API
  rejects malformed entries; safeConfigDTO exposes the field to the dashboard.

Implemented on feat/provider-cost-overlay (PR pending).