fix: teach agents the ToolSearch deferred-tool step (heavy-MCP sessions)

[azmym]

Root cause (reproduced empirically this time)

When many MCP servers are connected, Claude Code switches MCP tools into deferred mode: the tool name is shown in a system reminder but its schema is NOT loaded, so a direct call fails with an input-validation error until the caller materializes it via ToolSearch.

The five Gemini agents instructed themselves to "use the gemini_search_grounded MCP tool" but never mentioned ToolSearch. So a subagent in a heavy-MCP session could not find a directly-callable Gemini tool, "hunted for the tool name," burned its turn budget, and ended mid-investigation without emitting its JSON. The orchestrator then fell back to plain web search. This matches the user's reports exactly:

• the researcher "was still hunting for the gemini_search_grounded tool name"
• "the Gemini grounded-search tools aren't reachable from my toolset"

Evidence

Test	Result
Direct call to mcp__plugin_gemini-plugin_gemini__gemini_search_grounded (after ToolSearch)	✅ real answer + citations — server is healthy
Real installed gemini-researcher on a German-law question	✅ grounded JSON, confidence: high
Generic subagent probe	❌ tool "NOT directly callable, appeared only in the deferred-tools list, schema not loaded"; ✅ became callable after ToolSearch
Disk audit	installed v0.5.0 agents have no tools: allowlist (v0.4.1 fix present), and none mention ToolSearch

This is a third, distinct cause, separate from the v0.4.1 tools:-allowlist bug and from the stale-session case gemini-doctor already covers. It is intermittent because a light session loads the tools directly, and a capable model sometimes rediscovers ToolSearch unaided (which is why it "sometimes works").

Fix

• All 5 agents: added a **Deferred tools (do this FIRST)** step to the "Tool availability (fail loud)" section — call ToolSearch with a namespace-agnostic keyword query (e.g. gemini search grounded), then call the exact tool name returned; treat the tool as missing ONLY after ToolSearch returns no Gemini match. Keyword queries match both install namespaces and contain no mcp__ token, so the existing no-hardcoded-namespace test stays green.
• gemini-consult skill: documents the deferred-tool path and routes to /gemini-plugin:gemini-doctor when an agent returns unavailable/unknown with a missing-tool error.
• tests/mcp-namespace.bats: two new guards — every agent must mention ToolSearch, and must use a keyword query (not a hardcoded mcp__ path).
• Bumped to 0.5.1; CHANGELOG updated.

Verification

bats tests/ → 83 passed, 0 failed (added #72, #73). No em/en dashes in the diff.

Release

After merge, tag v0.5.1 to let the new release workflow (#19) publish it.