tracking(runtime): finish provider-native web search

[Phoenix500526]

Problem

#2152 established provider-hosted WebSearch for OpenAI Responses and Anthropic
Messages, but the remaining provider wires and verification gaps are split
between documentation and isolated issues or pull requests. There is no single
place to track ordering, shared invariants, and completion.

This is a coordination tracker. Existing provider-specific work remains
authoritative and should be linked rather than reopened.

Desired outcome

Every provider in the capability matrix has an explicit, fail-closed status.
Supported adapters execute search in the primary model request and preserve
provider events, citations, durable replay data, credential isolation,
privacy-mode removal, scoped tool ceilings, and explicit no-fallback behavior.

Checklist

• Reconcile or supersede the stale effective-wire follow-up fix(search): keep native routing aligned with model wire #2179
• Complete Alibaba/Qwen through feat(runtime): map Alibaba Token Plan native Harness tools into Maka #3163 and feat(runtime): add Qwen3.8-Max native WebSearch #3685; do not reopen duplicate feat(runtime): add Alibaba native WebSearch for Qwen3.8-Max #3680
• Reconcile DeepSeek restoration across feat(runtime): restore DeepSeek native WebSearch on Open Responses #3689, feat(runtime): restore DeepSeek native WebSearch #3694, and feat(runtime): adopt Open Responses extension codecs for DeepSeek tools #4107
• Live-verify the existing MiniMax Anthropic-compatible route
• Add opt-in Anthropic web_search_20260209; retain 20250305 as baseline
• Add Gemini grounding with Google Search
• Add the Z.AI native model tool; exclude its standalone Search API
• Add the OpenRouter web plugin
• Define cost/result-row contracts, then add Mistral Agents and Groq Compound
• Gate every adapter by exact provider/model/wire and cover lowering, streamed results, citations, replay, privacy, tool ceilings, and no fallback

Non-goals

• Replacing Tavily or adding silent fallback
• Inferring hosted search from generic protocol compatibility
• Adding Z.AI's standalone Search API to the model-native path
• Reopening work already owned by provider-specific issues or PRs
• Provider-native Tool Search for large catalogs ([RFC] Provider-native Tool Search for large tool catalogs #1382)

Documentation

• Provider-hosted web search capability