Port upstream 0.55.0: optimize OpenCodex pricing

[Finesssee]

Summary

• port upstream CodexBar v0.55.0 OpenCodex pricing hot-path optimization from Price OpenCodex usage once per entry and stop per-entry catalog/overlay reloads steipete/CodexBar#3136
• load the cached models.dev catalog once per OpenCodex aggregate and pass an immutable pricing snapshot through row pricing
• preserve historical request-day pricing and existing custom-pricing precedence
• avoid per-entry models.dev cache metadata checks while keeping output semantics unchanged

Upstream evidence

• target release: v0.55.0
• upstream commit: d8d03a122 / Price OpenCodex usage once per entry and stop per-entry catalog/overlay reloads steipete/CodexBar#3136

Validation

• cargo test -p codexbar opencodex --no-fail-fast -> 6 passed
• focused rustfmt on touched Rust files
• git diff --check -> pass

Scope

Isolated OpenCodex spend performance port. No provider-auth, UI, or version bump changes.

[Finesssee]

Thermo-nuclear review

Finding (structural/performance blocker): pricing_snapshot() deep-clones the entire models.dev catalog. ModelsDevCache::load() already returns an Arc<ModelsDevCacheArtifact>, but the new snapshot immediately does artifact.catalog.clone(), copying the full provider/model HashMap tree once per aggregate. That replaces repeated metadata checks with a potentially large ownership copy.

The code-judo version is to let ModelsDevPricingSnapshot retain shared immutable ownership, e.g. an Arc<ModelsDevCacheArtifact> or Arc<ModelsDevCatalog>, and perform lookups through it. That keeps the bulk-pricing context genuinely cheap and avoids widening a central pricing boundary around copied state.

Verdict: NOT APPROVED.