feat: smart tool shortlisting per turn

[Shifat7]

What changed

Before buildRelayPrompt(), each turn is classified locally into one of final_answer_only / read_or_search / edit_code / run_command / debug_failure / unknown, and only the client tool categories that task type plausibly needs are forwarded to Hyperagent. Function tool schemas are minimised to argument names, types, and required lists; descriptions are capped (160 chars default).

• New exports in src/protocol.mjs: classifyTurn(input), shortlistTools(tools, taskType, config); extractClientTools applies them when enabled.
• Config: enableSmartToolSelection (default true), forwardFullToolSchemas (default false), maxToolDescriptionChars (default 160). Old behaviour available via enableSmartToolSelection: false.

Why it saves credits

Tool schemas are a large fixed cost on every request of every tool loop.

Measured locally (synthetic 32-tool toolbox, request "Run the test suite."):

Before: 32 tools forwarded, 10,431 tool-schema chars
After:   4 tools forwarded,  1,127 tool-schema chars

Safety rules (all test-enforced)

• Exact tool names preserved; nothing invented.
• Never removes all tools when the user asked to inspect/edit/run/patch; edit and debug tasks keep patch + read + shell chains.
• Conversations that already used tools are never reduced to zero.
• Unknown turns keep the full inventory; disabling the flag restores byte-for-byte prior behaviour.

Risks/caveats

A mis-classified first turn can omit a specialised MCP tool for that one turn; the relay falls back to plain text naming the missing tool (existing parseRelayOutput behaviour) instead of executing it. Rephrase or disable if that bites.

Testing

npm test: 82 pass, 0 fail (3 pre-existing skips). New tests cover classification taxonomy, per-task shortlists, edit/debug chain preservation, mid-task retention, disable/unknown pass-through, and schema minimisation shape.