refactor(rust): move implementation to rust/clawcode workspace

[huagusam]

Restructures the Rust codebase into a self-contained \clawcode\ workspace under
ust/clawcode, collocating claw config (agents, skills) with the Rust crates.

• Move crates: agents, api, claw-cli, commands, compat-harness, migrate-patch-names, mock-anthropic-service, plugin-types, plugins, runtime, telemetry, tools
• Drop the old flat
  ust/\ crates tree
• Add
  ust/clawcode/claw/\ workspace config with agents and skills
• Gitignore local .env\ secrets (\claw/.env, \claw/.env_deepseek)

Verified: \cargo build --release\ succeeds in the new workspace.

[1716775457damn]

This restructure looks good — collocating the claw config (agents/skills) with the Rust crates under rust/clawcode keeps configuration in sync with the code, and gitignoring the local .env secrets is the right call. Since cargo build --release passes in the new workspace, the move seems safe. Thanks for cleaning up the old flat crates tree too.

[1716775457damn]

Looking at the reasoning-effort registry in 93ca225: the per-provider level/wire mapping plus fail-fast preflight validation is the right shape, and omitting reasoning_effort entirely for off (rather than sending "off") is correct — several OpenAI-compatible servers reject an explicit off value.

Two edge cases I don't see covered:

1. Budget vs max_tokens: max maps to a 32000 thinking budget, but Anthropic rejects requests where the thinking budget is >= max_tokens. Preflight validates that the level is supported, but as far as I can tell it doesn't compare the resolved budget against the configured max_tokens, so a --reasoning-effort max run with a small --max-tokens would fail at request time instead of up front — which is exactly the class of failure the preflight check exists to prevent.

2. Capability detection on self-hosted backends: fail-fast depends on knowing whether the endpoint supports reasoning at all. For a self-hosted Anthropic-compatible server (the scenario fix: harden the agent loop against a self-hosted Anthropic-compatible backend #3292 is hardening against), the capability table may be incomplete or the proxy may silently strip the field — then a valid-but-unsupported level either fails spuriously or gets dropped without any signal. Worth a way to mark a provider as reasoning-unknown and degrade to a warning rather than a hard failure.

Neither is a blocker for the restructure itself; the workspace move and the .env templating look good.