feat(acp): honor managed-agent reasoning effort

[rsaulo]

Problem

Buzz Desktop exposes an effort selector (EffortSelectField), but that field is not rendered for BYOH harnesses; I verified this in the UI with several models. Separately, BUZZ_AGENT_THINKING_EFFORT has exactly one runtime consumer today, in crates/buzz-agent/src/config.rs; nothing in buzz-acp reads it. As a result, managed agents running Claude Code, Codex, or opencode stay at the harness default with no effort control.

Why named effort through _meta, not MAX_THINKING_TOKENS

A token budget loses operator intent: high, xhigh, and max all collapse to 32,768 tokens. The Claude Agent SDK exposes named effort, and the adapter spreads ...userProvidedOptions after its defaults, so the client-provided value wins while preserving those distinctions.

Why the model-family table lives in buzz-core

This is a means, not an end: both the native runner and the external ACP path need the same Anthropic model-family classification. Sharing it avoids two tables silently diverging when a new model family is added. buzz-agent behavior remains identical, and its existing suite remains intact.

Harness-specific behavior

• Claude Code: sends named effort through _meta.claudeCode.options (and retains manual token budgets for legacy model families).
• Codex: deep-merges the effort into CODEX_CONFIG, preserving unrelated settings such as approval_policy and sandbox_mode.
• opencode: deliberately does not write an operator config file. A configured effort fails explicitly at startup with guidance instead.
• A context-window suffix such as [1m] is ignored for model-family classification.
• An unversioned alias omits the option with an actionable message, because it cannot safely be classified as adaptive thinking versus a manual token budget.

Tests

• cargo test -p buzz-core
• cargo test -p buzz-agent
• cargo test -p buzz-acp

[rsaulo]

Closing this in favour of #4625, and of #4557 which landed the ACP-side wire path.

Both solve the problem this PR was aimed at, and solve it better. The thing I got wrong here was hardcoding an effort axis: #4625 routes the effective value through BUZZ_ACP_EFFORT_LEVEL and lets the adapter's advertised thought_level option define the vocabulary, so Claude, Codex and OpenCode each get exactly the levels their current model supports — discovered rather than tabulated. The single-destination-key projection and the suppress set also close a class of double-authority bug this PR never addressed, and applying against the post-switch snapshot gets the model-then-effort ordering right. My Anthropic model-family table in buzz-core would have drifted within a release or two.

One narrow gap I'd like to file separately: models exposing a manual thinking budget rather than named effort (claude-3*, claude-opus-4-5) advertise no thought_level option, so a configured effort is silently dropped for them. Happy to open an issue.

Thanks for the thorough design here.