Fail the turn instead of dispatching truncated tool calls

[TheGreatAxios]

Summary

Fixes CL-7783: a truncated Anthropic tool_use (max_tokens cut off mid-JSON) was dispatched as a well-formed call — {"_raw": ...} args or silently dropped params — and reached the reactor as a normal tool call. Now the adapter surfaces stop_reason on the turn and end-of-stream finalization fails the turn with an actionable error instead of dispatching unparseable args.

Changes

• Adapter (vendor/intx-inference/src/providers/anthropic.ts): parse stop_reason in MessageDelta, plumb through LastUsage/streamState, emit on inference.usage (arktype + static union in vendor/intx-types/src/runtime.ts).
• Harness (vendor/intx-inference/src/harness.ts): finalizeStreamTurn collects parse failures across all three paths and throws a HarnessError naming the tool and failure kind (truncation vs malformed) instead of dispatching. No _raw fallback remains; the dead _raw consumer in src/plugins/path-escape-plugin.ts is removed.
• Reactor: truncated tool calls surface as retryable turn failures, never as tool results.
• Ledgers: local vendor patch recorded as stopgap + promotion candidate in vendor/intx-inference/PATCHES.md and vendor/intx-types/PATCHES.md; docs/VENDORING.md updated. Upstream untouched.

Verification

• New regression test drives the exact CL-7783 event sequence (content_block_start tool_use Bash, partial input_json_delta, message_delta stop_reason: max_tokens, message_stop) and asserts no tool_call block reaches the reactor: bun test vendor/intx-inference/src/providers/anthropic.test.ts — 7 pass.
• Targeted suites (inference-harness.test.ts, agent-tool-dispatch.test.ts) green; bun run check green.
• OpenAI-compatible adapter audited for the same path (see review comment / report).

Fixes CL-7783.

[linear-code]

CL-7783