Bump rig to 0.41 so reasoning shows on LocalAI (#745)

[yogthos]

Fixes #745.

rig 0.39 only deserialized delta.reasoning_content. Providers that stream reasoning as delta.reasoning — LocalAI, and as it turns out Cerebras — had it dropped by serde, while text and tool calls carried in the same delta worked fine. That's why the reporter saw an empty reasoning panel and normal output otherwise. Parsing the log they attached: 196 chunks, 188 with reasoning, zero with reasoning_content.

rig 0.40 added reasoning_content.or_else(|| reasoning) upstream. Checked against opencode, which the reporter noted does not have the bug: @ai-sdk/openai-compatible does delta.reasoning_content ?? delta.reasoning. Same precedence, so this matches the reference implementation.

The bump

0.41 splits the classic runtime into rig-agent behind an agent feature, replaces Tool::definition with description/parameters across all 37 tools, drops ToolDyn for a concrete struct, and makes Agent::preamble and Agent::model private.

• Tools move to the context-free PortableTool contract. The loop has driven its own dispatch since 4.5h-6 and never used rig's ToolContext.
• The erased-tool seam is now dirge's own DynTool. It has broken on two consecutive rig releases and does not need to be rig's. 0.39's semantics are ported exactly, including the null -> {} fallback for all-optional args and the rule that a String output is not double-quoted.
• AnyAgentInner holds the completion model directly instead of a rig Agent it only ever read the model back out of.

Cerebras fallout

Capturing the reasoning made Cerebras 400 on the next turn: property 'messages.N.assistant.reasoning_content' is unsupported. Confirmed a real regression by running the smoke test on 0.39 (passes) and 0.41 (fails), not just assuming a flaky live API.

The field is renamed to reasoning at the wire boundary rather than dropped, so the model keeps its own reasoning in context. That is what @ai-sdk/cerebras does for opencode. Backends that want reasoning_content are untouched — DeepSeek needs it on tool-call turns, and llama.cpp/LocalAI chat templates read it back out of the assistant turn. Only the OpenAI Responses API still has the block dropped, since it keys reasoning to encrypted ids we don't retain.

Dependencies

rmcp was on three versions at once: ours, one from rig's rmcp feature (which nothing here used), and one from agent-client-protocol. That feature is gone, acp goes to 2.0 and rmcp to 3.1 — one copy, at latest. rusqlite, sysinfo, sha2, jsonschema, base64, compact_str, http and notify-rust go current too. tree-sitter stays at 0.25; 0.26 does not resolve against the grammar crates.

Verification

• cargo fmt --all clean
• RUSTFLAGS="-D warnings" cargo clippy --all-targets exit 0
• cargo nextest run --bin dirge --retries 2 --no-fail-fast — 4995 passed, 1 skipped, 0 failed (baseline 4960)

The live h7_cerebras_tool_dispatch_completes_round_trip passes, so the rename is verified against the real API rather than only in unit tests.

Two things worth a reviewer's eye:

• The migration reindented raw string literals in three tool descriptions (memory, session_search, spec), silently changing model-visible prompt text. Restored, and every changed file is now checked so no long or multi-line literal value differs from HEAD.
• My first cut of the Cerebras rewrite removed reasoning_content before checking whether reasoning was already present, which lost the field on a skipped message when a later one triggered the re-serialize. Fixed, with a test that fails against the buggy version.