fix(providers): dispatch opencode-go qwen3.8-max unchanged

[gonisulaimann]

OpenCode Go ships the GA qwen3.8-max upstream, but the global deprecation map in modelDeprecation.ts still forwards it to qwen3.8-max-preview. That mapping predates the GA model (its own comment says to drop it once Alibaba ships a distinct qwen3.8-max — that condition is now met), so every request for opencode-go/qwen3.8-max was rewritten and rejected upstream with 401 Model qwen3.8-max-preview is not supported.

What changed:

• open-sse/services/modelDeprecation.ts — removed the stale qwen3.8-max → qwen3.8-max-preview mapping.
• src/shared/constants/modelSpecs.ts — qwen3.8-max gets its own spec with the same 1M-context envelope, so the context-preflight coverage the rewrite used to provide is preserved through the spec itself; the preview spec no longer aliases the GA id.
• open-sse/config/providers/registry/opencode/go/index.ts — registered the GA row with the qwen routing flags (targetFormat: "claude", supportsReasoning: true, supportsVision: false) so hasKnownProviderModel also holds for the direct path.

Providers that legitimately expose the preview id are untouched: qwen3.8-max-preview still resolves and dispatches on qoder and bailian-coding-plan.

Tests (tests/unit/qwen3-8-max-ga-14181.test.ts, written first — all four failed on the base tip):

• resolveModelAlias("qwen3.8-max") and with the opencode-go provider both return the id unchanged.
• qwen3.8-max-preview still resolves on qoder and bailian-coding-plan, and both registries keep the preview row.
• opencode-go declares qwen3.8-max with the routing flags above.
• qwen3.8-max has its own MODEL_SPECS entry and the preview spec no longer aliases the GA id.

Results:

• node --import tsx/esm --test tests/unit/qwen3-8-max-ga-14181.test.ts — 4/4 pass.
• Related suites (model-deprecation, model-deprecation-aliases-11503, resolve-model-alias-index-8697, bailian-coding-plan-provider, qwen-cloud-token-plan-discovery, model-capabilities-registry) — 135/135 pass.
• tsc --noEmit clean on the touched files; eslint clean; prettier clean.

Fixes #14181

[gonisulaimann]

Closing in favor of #14242, which fixes the same issue and does it more correctly.

When I re-verified the registries at the current tip, the ground truth is: qoder and bailian-coding-plan serve only qwen3.8-max-preview, while qwen-cloud-token-plan, alibaba, qwen-cloud, kilocode, clinepass and xkiro list the GA id. My branch deleted the global mapping outright - which fixes opencode-go but would regress bare-id qwen3.8-max requests against qoder and bailian. #14242 keeps the rewrite for the preview-only providers and lets the provider-aware exemption from #11503 pass the GA id through on opencode-go, which is the mechanism the repo already built for exactly this. Its test suite also pins the qoder rewrite staying intact, which mine asserted away.

The parts of my branch that were right (the GA MODEL_SPECS row, the opencode-go registry entry) are both in #14242, so nothing is lost here.