feat(provider-opencode-go): add OpenCode Go Chat Completions provider worker

[faramirezs]

What

New Rust worker provider-opencode-go: an LLM provider worker behind
llm-router speaking the OpenCode Go API — Chat Completions
(https://opencode.ai/zen/go/v1/chat/completions), SSE streaming, live model
discovery, auth/error taxonomy, reasoning-effort mapping, tool calling, and
structured output. Registers provider::opencode_go::stream / refresh_models
/ abort with the router, binds identity via registration token (state scope
provider-opencode-go), and reads OPENCODE_GO_API_KEY as credential.

Why

The stack has no way to route chat completions to an OpenCode Go subscription.
The OpenCode Go API is Chat Completions compatible, so the existing
provider-openai worker ports directly — same protocol, same relay/pump
scaffold, same error taxonomy, with the OpenAI-only surfaces (embeddings,
Responses API) dropped as dead code.

How it works

• Credentials: resolved per request via router::provider::resolve
  (config slice → OPENCODE_GO_API_KEY env on the router → none); sent as
  Authorization: Bearer.
• Wire / streaming: Chat Completions SSE → AssistantMessageEvent frames
  into a router-owned channel; ping ≥ every 30s of silence; a failed channel
  write (router::abort / caller gone) drops the SSE receiver and aborts the
  in-flight HTTP request. Stream-path tracing::debug for provider-side
  observability.
• Catalog / discovery: live GET /v1/models supplies bare ids; each is
  enriched from a hardcoded curated metadata table (src/curated.rs) prepared
  from models.dev (2026-08-03) — context window, reasoning support/effort
  levels, tool-call and structured-output capability for the maintainer's
  curated model set (24 models.dev entries + hy3-preview). Ids outside the
  table keep conservative defaults (128K, no thinking, tools on) — same
  pattern as provider-openai's curated.rs.
• Reasoning: thinking_level maps to the upstream reasoning_effort only
  when the model's curated effort list accepts the level (e.g. grok-4.5
  accepts low/medium/high; deepseek-v4-flash and glm-5.2 accept
  high/max; hy3 accepts none/low/high); toggle-only models and
  unknown ids stream without the field.
• Registration: self-declares via router::provider::register with
  backoff until acked, re-declares on router::ready; registration_token
  persisted in iii-state (scope provider-opencode-go).

Scope / caveat

• Port of provider-openai: mechanical renames only in most files; the
  OpenAI-only surface (ApiMode, embeddings, Responses-API event handlers and
  thinking deltas, curated reasoning-fallback ladder, luna guard) is dropped —
  OpenCode Go has no such surface.
• Not a replacement or rewrite of the Node opencode/ worker —
  different role (CLI wrapper vs provider), no overlap; both install side by
  side.
• hy3-preview is listed by the live GET /v1/models but the chat endpoint
  currently returns ModelNotFound — an upstream inconsistency; the curated
  row keeps conservative defaults and the provider surfaces the upstream error
  cleanly.
• External contributor: no Linear ticket — no-ticket label applied.

Repo wiring

• Root README.md Modules row added for provider-opencode-go
  (alphabetical, between provider-openai and provider-xai)
• Release wiring: create-tag.yml, release.yml, alpha-release.yml, and
  .github/scripts/discover_changed_workers.py
• llm-router/README.md reference note (same structure as provider-openai)

Verification

• cargo fmt --check and cargo clippy --all-targets --all-features -- -D warnings clean
• cargo test --all-features: 67 pass — 57 lib unit + 2 bin unit + 4
  schema/golden + 4 integration
• Integration tests run against the real engine (v0.22.0) + llm-router with a
  stubbed upstream: chat stream end-to-end (incl. cache_read usage),
  401 → auth_expired error frame, refresh_models → catalog from the
  curated table, re-declare on router::ready
• cargo build --release OK
• Live smoke (real engine + llm-router 1.4.0 + OpenCode Go subscription):
  refresh_models → 25 models; router::models::list → 25 curated entries
  with metadata; router::complete served on glm-5, deepseek-v4-flash
  (incl. thinking_level: high → reasoning_effort), and hy3

Test plan

• cargo fmt --check
• cargo clippy --all-targets --all-features -- -D warnings
• cargo test — 67 pass
• III_ENGINE_BIN=$(which iii) cargo test --test integration -- --test-threads=1
• Live: iii worker add provider-opencode-go → refresh_models →
  router::models::list → router::complete

[github-actions]

skill-check — worker

0 verified, 54 skipped (no docs/).

Layer	Result
structure	✓
vale	✓
render	✓

Three for three. Nicely done.