[KytranKatarn]

Changes

• Explicitly forward input.model to AI Horde. Earlier Model None not known failures cannot be attributed to offline workers when the requested model was not sent.
• Add a provider-scoped instruction to emit the actual tool envelope rather than narration. Shared prompts for other executors remain unchanged.
• Opt AI Horde into strict parser mode: exact request nonce and requested tool name are mandatory. Missing/wrong nonce, unknown names and incomplete blocks remain non-executable text. Other executors retain compatibility defaults.
• Preserve caller max_tokens; short truncated output is not proof of model noncompliance.

Verification

• 58/58 targeted executor/shared-parser regression tests pass (JSON, SSE, no-tools, nonce/name rejection and truncation).
• Four targeted Reject invalid output token budgets cheaper-inference/OmniRoute#7379 adapter usage tests also pass; this is NOT fresh live usage proof.
• Synthetic outgoing body captured with nonce redacted in the session artifact. No real tool execution performed.

Limitations

Live model-compliance testing is blocked by the platform OmniRoute provider dispatchable=false. No routing policy was enabled or bypassed. This change is not deployed.

[KytranKatarn]

Added the remaining live-only AI Horde emulated free-pool opt-in to prevent an integrated rebuild removing that route. Total now 4 files; 59/59 targeted tests pass. Also captured the synthetic upstream body (request nonce redacted); no live inference or client tool execution performed. Pricing PR #21 is merged, but its build is held until this provenance gap is reconciled.