refactor: unify the two Anthropic-likeness classifiers (hoist vs cache-control)

[anandgupta42]

Found during v0.8.3 release review (CTO persona). Deferred because it touches a hot path and changes cache-control behavior for gateway traffic — needs design + careful testing (>30 min), not for a patch release.

Context

There are now two independent predicates deciding whether a model is "Anthropic-like":

1. isAnthropicLikeModel in packages/opencode/src/session/prompt.ts — gates the <system-reminder> system-role hoist (added in fix: plan-mode refusals on altimate-default + warning false-positive (#887) #888). Built on familyVendor + providerID + api.id/npm heuristics.
2. The inline cache-control check in packages/opencode/src/provider/transform.ts:284-295 — gates Anthropic cache_control injection. Hard-codes model.providerID === "altimate-backend" as Anthropic-for-caching.

These deliberately diverge today: transform.ts treats the altimate-backend gateway as Anthropic-for-caching, while isAnthropicLikeModel classifies the same gateway (family: openai) as non-Anthropic. The divergence is correct for the #888 fix, but the two classifiers will drift further apart as gateway models expand, and a future maintainer could unify them wrongly.

Suggested fix

Extract a single shared predicate (built on familyVendor) used by both the hoist decision and the cache-control decision, with explicit handling of the gateway case so the two behaviors can't silently drift. Verify cache-hit rates on gateway traffic don't regress.

Scope

packages/opencode/src/session/prompt.ts, packages/opencode/src/provider/transform.ts, packages/opencode/src/provider/family.ts.