[Contract]: Reconcile Vela span/set and truncation semantics before native parity #169

Parent and existing work

Contract/acceptance follow-up for #137 and #61. Do not duplicate #151, the existing Vela 0.3B reference-worker PR by MrDongsls, or #152, the configurable encoder / ModernBERT work. The first reference and native foundation already have implementation carriers.

At the inspected #151 head 8f3ed1b7f41e7fde47b2dc6cccf3ea0f62281d8f, its body deliberately forwards all five question kinds, including span/set and top-level spans/sets/thresholds, and follows upstream state truncation. The older #137 proposal instead described rejecting span/set and no silent truncation. #61 has discussion of reference-defined fields but does not yet record these Vela-specific decisions as an accepted common contract.

A fresh GitHub read on 2026-10-25 reports #151 and #152 (8d7be88ac271d4a983ef46fe2999dc31a71a6861) as mergeable: CONFLICTING. Coordinate integration on the existing carriers and rerun their acceptance at the resolved head; open PRs and historical passing checks are not merged or current-head support.

Decision to record

1. Whether the shared transport permits these model-defined extra response fields unchanged. Keep the Rust frontend a byte-forwarding transport; do not silently reinterpret span/set as exclusive Choice.
2. Whether Vela's upstream state truncation is the accepted model-specific behavior, with explicit documentation of its guardrail consequences, or an intentionally stricter worker contract. Do not change it incidentally during native porting.
3. Preserve Vela's reference confidence statistics and single-rendered-sequence input-token count; do not replace them with another worker's entropy or per-question sum. Keep learned temperature tau separate from inference calibration T.

Acceptance fixtures for the selected policy

• Record the decision in [RFC]: Write down which parts of the /v1/systemone contract all workers share #61's contract documentation and reconcile [New Model]: Support the Vela 2.0 series #137's acceptance wording with Vela: serve Vela 2.0 0.3B through a Python reference worker #151's final supported scope.
• Span cases retain the named state field, router/broad-head identity, label, probability and Unicode-code-point offsets. Include CJK and non-BMP characters so byte/UTF-16 offsets cannot pass.
• Set cases retain independent per-label probabilities, per-request thresholds and multi-label selection. No cross-label softmax or forced single selection.
• Preserve Noul views and the corresponding top-level extension fields together; check direct worker/frontend byte or semantic parity under the documented timing-field policy.
• Cover invalid span/set requests, error precedence, oversized schema versus oversized state, and the actual upstream truncation boundary.
• Use the pinned 0.3B CPU FP32 oracle before native comparison. Report native precision/backend tolerances separately.

Suggested non-overlapping contribution for psv666: contract acceptance and additional native parity cases after the reference/oracle is available. MrDongsls remains the author of the existing reference/ModernBERT PRs. This proposal does not assign another contributor new work or claim native parity has passed.