fix(api): reconcile custom-provider catalog identities for client enrichment

Environment

• OmniRoute 3.8.52, image digest sha256:90f903d7e9245973dc23f7f0a3afc0c63502c9bd25961097843361689df6474b
• Observed through the OpenCode v2 client plugin @omniroute/opencode-plugin-v2@0.1.0, but the defect is gateway-side (endpoint inconsistency), not plugin-side.

Summary

For custom openai-compatible-chat-* connections the enrichment overlay (/api/pricing/models) cannot be joined to the served catalog (/v1/models), for two independent reasons:

1. Key mismatch. Enrichment entries are keyed by the connection's internal provider string (e.g. openai-compatible-chat-8d3c46c4-20bc-40ad-a837-c475f578dc8a), while the served models carry routing prefixes (ih/..., owned_by: "ih"). Built-in providers work only because their key equals owned_by (e.g. kilocode).
2. Id-shape mismatch / stale sync. Per connection the model-id shapes disagree between the two endpoints:
   • connection ...8d3c46c4... serves ih/** on /v1/models (606 models, owned_by: ih) but its pricing entry lists ag/** ids (384 models, 206 prefixed, 0 ih/) - this looks like a stale sync from a previous prefix.
   • connection ...e7cc3ed0... serves prefixed si/** on /v1/models (520 models) but its pricing entry lists 449 bare ids.
   • The pricing catalog contains no si and no ih key at all.

Consequence

• lookupEnrichment() misses for every model of these connections: no providerDisplayName, no providerAlias, no pricing, no freeType.
• Client catalogs therefore show raw ids (si/kimi-k3, or bare kimi-k3 for unprefixed serves) with no provider attribution and cost 0/0 - even though owned_by is set correctly on all 4,497 served models.
• The free-tier summary reports large unmatched counts; plugin refresh log: catalog refresh: served=4659 default=59 presets=59 published=61 unmatched-free=471.
• /api/providers exposes neither prefix nor a prefix->connection mapping, so clients cannot repair the join locally from public endpoints.

Evidence

• /v1/models owned_by histogram: kilocode 1446, tokenrouter 750, ih 606, si 520, zenmux 396, aihorde 219, ...
• /api/pricing/models keys include openai-compatible-chat-8d3c46c4-... (first ids: ag/claude-opus-4-6-thinking, ag/claude-opus-5-5, ...) and openai-compatible-chat-e7cc3ed0-... (first ids: openai-gpt-oss-120b, minimax-m2.1, ...); no si, no ih key.
• Working counter-example: key kilocode with kilocode/... / kc/... ids joins fine and yields labels + pricing in clients.

Expected

• Key (or alias) enrichment entries by routing prefix / owned_by in addition to the connection id, or expose the prefix <-> connection mapping on /api/providers.
• Keep id shapes consistent between /v1/models and /api/pricing/models per connection; re-sync the pricing catalog when a connection prefix changes (the ag/ list of ...8d3c46c4... appears to be a leftover of an earlier prefix).

Workaround in place (client-side, local only)

The OpenCode plugin was locally patched to fall back to a provider label derived from owned_by when the enrichment lookup misses. Pricing and free markers remain unavailable for these connections until the gateway side is fixed.