feat(v1): rlm harness builtin_tools passthrough

[hallerite]

Companion to nano-rlm's contract-tooling change (PrimeIntellect-ai/nano-rlm#164), which moves builtin tool selection out of RLM_TOOLING/RLM_BUILTIN_TOOLS env vars and into the ai.prime.rlm/runtime-v1 contract.

• RLMHarnessConfig.builtin_tools: list[BuiltinTool] | None — the contract's builtin_tools. None (default) keeps rlm's ipython-only tool set and stays off the wire, so nano-rlm pins that predate the field keep validating; setting it requires a pin that understands it (nano-rlm#164 or later).
• BuiltinSkill synced with nano-rlm main's builtins (bash, fetch added — the fetch skill+tool merged in nano-rlm#132).
• disabled_tools rejection message updated (the tool set is no longer described as fixed).

Motivation: the fetch builtin tool is otherwise only reachable by smuggling env through harness.env, exactly the pattern nano-rlm#160 removes everywhere else. This also makes a clean fetch-as-skill vs fetch-as-tool A/B expressible purely in config.

Note

Add builtin_tools passthrough field to RLMHarnessConfig and expand BuiltinSkill set

• Adds a new optional builtin_tools: list[BuiltinTool] | None field to RLMHarnessConfig; when set, it is included in the RLM runtime metadata payload, and when None the key is omitted for backward compatibility with older pins.
• Introduces the BuiltinTool Literal type (bash, edit, fetch, ipython) and expands BuiltinSkill from edit/search to also include bash and fetch.
• Updates the reject_disabled_tools validator error message to point users at builtin_tools and builtin_skills instead of claiming a fixed tool set.
• Risk: configs that previously relied on the old fixed-tool-set error message wording will see a changed message; existing payloads are unaffected unless builtin_tools is explicitly set.

^{Macroscope summarized f6a75fb.}