feat(Research): add provider-specific execution guides for researcher agents

[Kharsys]

Summary

Three of the four Researcher agents (Gemini, Grok, Perplexity) currently have no provider-specific call code anywhere in the Research skill or in the agent files. With no API/CLI wire-format defined, each non-Claude agent collapses to the same backend (built-in WebSearch under a different persona name), so "multi-model research" via Standard / Extensive mode does not actually exercise multiple models. This PR adds the missing execution layer.

What's added

A new directory Releases/v5.0.0/.claude/skills/Research/Workflows/Execution/ with four guides, one per researcher:

• ClaudeExecution.md documents the WebSearch path (included for parity, no behavior change)
• GeminiExecution.md documents the gemini CLI invocation with Google Search grounding
• GrokExecution.md documents the xAI Chat Completions API call with live search enabled
• PerplexityExecution.md documents the Perplexity Sonar Chat Completions API call

Each guide owns the provider call shape (CLI or curl with jq -n --arg for safe user-input escaping), the model selection table mapped to research modes, citation extraction rules, URL verification via curl, the file-based output JSON schema, and a fallback policy that retries once at a lower model tier then writes a stub error file. Never silently swaps providers.

Agent changes

Each of ClaudeResearcher.md, GeminiResearcher.md, GrokResearcher.md, PerplexityResearcher.md gains step 3 in its Mandatory Startup Sequence: read the sibling execution guide. Existing persona, voice, output format, Research Philosophy and Research Methodology sections are unchanged. Nothing was removed.

SKILL.md change

A new ### Agent Execution Guides subsection under ## Workflow Routing points to the four guides and states the no-silent-fallback policy at the skill level.

File-based output schema

Each Execution guide documents a shared JSON shape for ${research_dir}/<agent-slug>.json:

{
  "agent": "claude|gemini|grok|perplexity",
  "query": "...",
  "summary": ["..."],
  "sources": [{ "title": "...", "url": "...", "verified": true }],
  "confidence": 0.0,
  "full_findings": "..."
}

The schema is shared across all four researchers so a downstream synthesis step can cross-reference findings agnostic of which provider produced them. Current upstream Research workflows synthesize inline as text rather than from files; this schema is therefore forward-looking with respect to the merged tree, standardizing the contract for any future file-based synthesizer. If research_dir is not provided by the caller, each researcher returns the same payload inline and skips file writing.

Correctness property preserved

A researcher whose provider call fails writes a stub file (summary: [], error reason in full_findings) rather than substituting WebSearch. The orchestrator's decision stays explicit: it can surface the gap to the user or substitute a different researcher. Silent substitution would defeat the purpose of multi-model research.

Why no OpenAI / GPT researcher

GPT plus web search via the OpenAI Responses API is architecturally similar to the Claude plus WebSearch path (LLM-with-search, citation-after-generation). It does not add an orthogonal coverage angle that the existing four researchers do not already cover. Grok contributes live X data and STEM strength; Perplexity contributes retrieval-first architecture with citations bundled into the response; Gemini contributes Google Search grounding. Each is structurally different. If a fifth researcher pointed at GPT-4o/5 is desired later, it can be added with the same Execution-guide pattern.

Smoke test recipe

Local setup expects ~/.claude/.env with XAI_API_KEY, PERPLEXITY_API_KEY, and GEMINI_API_KEY populated, plus the gemini CLI installed (validated against v0.39.1).

# Gemini
gemini -m gemini-2.5-flash -p "what is Bun JavaScript runtime" --grounding google_search

# xAI Grok
source ~/.claude/.env
BODY=$(jq -n --arg q "what is Bun JavaScript runtime" \
  '{model:"grok-3", messages:[{role:"system",content:"Research assistant. Cite sources."},{role:"user",content:$q}], search_parameters:{mode:"on"}, max_tokens:512}')
curl -sS -H "Authorization: Bearer $XAI_API_KEY" -H "Content-Type: application/json" \
  -d "$BODY" https://api.x.ai/v1/chat/completions | jq '.choices[0].message.content, .citations'

# Perplexity Sonar
BODY=$(jq -n --arg q "what is Bun JavaScript runtime" \
  '{model:"sonar", messages:[{role:"system",content:"Research assistant. Cite sources."},{role:"user",content:$q}], max_tokens:512, return_citations:true}')
curl -sS -H "Authorization: Bearer $PERPLEXITY_API_KEY" -H "Content-Type: application/json" \
  -d "$BODY" https://api.perplexity.ai/chat/completions | jq '.choices[0].message.content, .citations'

Expected: response content plus at least one citation URL. Missing-key scenarios should produce auth errors that each agent maps to the stub error file documented in the corresponding execution guide.

Diff stat

9 files changed, 583 insertions(+), 4 deletions(-)