Make offline scorecard pricing provider-aware

[Hmbown]

Problem

The offline scorecard currently accepts a model ID and usage but no effective provider. It calls model-only pricing helpers, so the same model name can be assigned API pricing even when the recorded turn came from a route where dollar pricing is unavailable or materially different (for example OpenAI API versus ChatGPT/Codex OAuth).

Draft PR #4332 fixes provider-aware pricing for active TUI turns, compaction/purge, runtime-thread records, tool children, and sub-agent mailbox usage. The standalone offline scorecard remains model-only and is intentionally called out as outside that claim.

The original scorecard work came from @findshan in #3388. Preserve that contributor credit in any redesign or harvested patch.

Expected behavior

• Persist or ingest effective provider provenance alongside every scorecard turn.
• Use provider-aware pricing; unknown/legacy provenance fails closed as unpriced rather than guessing.
• Keep token/cache metrics available even when dollar cost is unavailable.
• Preserve backward compatibility for old scorecard inputs and clearly mark their price as unknown.
• Do not infer provider from model name.

Acceptance

• Scorecard input/records carry optional effective provider identity.
• Codex OAuth tokens aggregate with zero fabricated API dollar cost.
• Same-model, different-provider fixtures produce different truthful pricing results.
• Legacy model-only fixtures remain readable and explicitly unpriced.
• CLI output and JSON distinguish zero cost from unavailable pricing.

[richardchen874-sys]

The instruction not to infer provider from model name is exactly right. The same model label can imply very different pricing or even no reliable dollar pricing depending on whether it came through API, OAuth, or another route.

I would carry provider provenance as a first-class field beside model and usage, then let legacy records fail closed as unpriced while still preserving token/cache metrics. That avoids false precision and keeps older scorecards useful.

I am testing an OpenAI-compatible multi-model API layer around official Chinese models, and provider provenance is one of the first things I try to keep explicit. Is the offline scorecard used for billing decisions, or mainly for developer cost visibility?