Build orchestrator prompt sections conditionally on registered tools (semantic_search-less MCP config)

[vitali87]

Follow-up to #1199 / PR #1200, prompted by review feedback there.

Gap

#1200 makes the orchestrator system prompt reference the registered short tool names, closing the rename drift. But the prompt text itself is static: when a canonical tool is not registered, the prompt still instructs the model to use it (with a runtime warning logged). One legitimate configuration hits this today: the MCP path (codebase_rag/mcp/tools.py) omits semantic_search when the vector backend is unavailable, yet the built prompt still contains the whole "WHEN TO USE SEMANTIC SEARCH FIRST" strategy section (prompts.py, the t.semantic_search sites around lines 156-183) telling the model to start with a tool that is not in its schema.

For prose-following models this is the same failure mode #1199 documented — a call to an unregistered name is dropped silently and the run burns retries on empty turns — just narrower: it needs a config where the tool is genuinely absent, rather than affecting every install.

Work

Make the prompt sections conditional on tool availability rather than interpolation-only:

• extract_tool_names (or a sibling) should expose which canonical tools are actually registered, not just their names — e.g. return (ToolNames, frozenset[AgenticToolName]) or make the availability set a field.
• build_rag_orchestrator_prompt builds the semantic-search strategy subsection (2a/2c workflows) only when semantic_search is registered; the graph-first guidance in 2b becomes the unconditional default in its absence.
• Same treatment for any other tool that can legitimately be absent (structural_search/structural_replace are also conditional in the MCP path, though they are not in ToolNames today).
• The runtime warning from fix: derive orchestrator prompt tool names from AgenticToolName so the prompt matches the registered tools #1200 stays — it then indicates a genuinely unexpected registration gap rather than a known configuration.

Tests

Extend test_prompt_tool_names.py: build the prompt from a tool list without semantic_search and assert the prompt neither backtick-references it nor instructs starting with semantic search; with the full list, assert the strategy section is present (guards against the section silently disappearing).