[Improve] Require explicit LLM usage source attribution

[roomote-roomote]

​Opened on behalf of @mrubens. Follow up by mentioning @roomote-roomote or in the web UI.

What changed

• Audited every tracked llmUsage.record entry point, including the SDK wrapper, worker SDK usage, harness code, and generated sandbox plugin scripts.
• Made source required by the in-repo sdk.llmUsage.record() helper so future callers must identify their usage surface.
• Reject blank and whitespace-only source values at the tRPC boundary, and trim meaningful source values before recording them.
• Kept the tRPC input optional and preserved input.source ?? 'opencode' for backward compatibility with deployed clients that omit the field.
• Added contract and router tests covering required attribution, blank-value rejection, and normalization.

The audit found no current in-repo caller of llmUsage.record, so there were no existing missing callers to assign new source strings. Active recording paths already attribute usage explicitly: task inference uses opencode, the inference gateway uses inference-gateway, and non-task inference uses semantic values from NON_TASK_INFERENCE_SURFACES.

Why this change was made

The endpoint fallback can silently collapse newly added SDK callers into the generic opencode bucket, while explicit blank strings can create a separate unattributed bucket. Requiring attribution in the SDK and validating explicit values at the RPC boundary prevents both cases without breaking older deployed clients that still need the server fallback.

Impact

New TypeScript callers of the SDK helper must provide a source. Explicit blank sources are rejected, padded sources are normalized, and direct or older tRPC clients that omit source remain compatible.

[roomote-community]

No new code issues found. See task

• Reject blank and whitespace LLM usage sources (packages/sdk/src/llm-usage.ts:7).

_{Reviewed 26cc524}

[roomote-roomote]

Fixed blank and whitespace LLM usage source handling in 26cc524a. The RPC now rejects blank values, trims meaningful values, and retains omission compatibility.