Add canonical human-benchmark evaluation provenance

[NeoLorenzo]

Audit Priority: P1
Audit Type: FIX

Problem / Opportunity

tools/evaluate_phase1.py recognizes only full 250-map validation/test runs as canonical. A complete evaluation across all 17 maps in the human_benchmark pool is therefore labeled partial-smoke-noncanonical, even when it covers the entire official human-benchmark split under the intended protocol.

That classification is unsuitable for evidence used in the human-relative Phase 1 decision gate. A complete human-benchmark model run needs an explicit machine-readable protocol designation that distinguishes it from partial/smoke runs without conflating it with the separate validation/test canonical protocols.

Why This Matters

Issue #15 must be able to reject incomplete or nonconforming model evidence and accept a complete protocol-conforming human-benchmark run without relying on filename conventions or free-text interpretation. The current boolean canonical plus classification fields cannot express that distinction correctly.

Evidence

• tools/evaluate_phase1.py sets canonical = (pool in (validation, test) and len(maps) == 250 and ... full policy/repeat configuration ...).
• The evaluator then sets classification to canonical only when that boolean is true; every complete human_benchmark run is therefore emitted as partial-smoke-noncanonical.
• load_verified_pool(...) and the split manifest already expose the authoritative human_benchmark pool identity and map count, so complete-vs-partial membership can be verified rather than inferred.
• docs/human-benchmark.md treats the 17-map split as the official persistent benchmark and expects future model results to join against immutable human records.
• Issue Evaluate the active Phase 1 reference on the human-benchmark pool #13 owns running the active reference; this issue owns making the resulting protocol status machine-readable and semantically correct.

Proposed Outcome

Extend evaluation provenance so a complete protocol-conforming 17-map human-benchmark evaluation is explicitly recognized as the canonical model-side human-benchmark comparison protocol (or an equivalently unambiguous designation), while preserving the existing validation/test canonical semantics.

Definition of Done

• A full protocol-conforming 17-map human_benchmark evaluation is no longer labeled partial-smoke-noncanonical.
• Saved evaluation artifacts contain an explicit machine-readable protocol/classification identifying the run as the complete human-benchmark comparison protocol.
• Protocol recognition verifies complete membership of the authoritative 17-map human-benchmark split and the required model-side policy/repeat semantics rather than relying only on map count.
• Partial human-benchmark runs, --max-maps smoke tests, reduced/nonconforming policy sets, or otherwise incomplete runs remain clearly noncanonical for the human-benchmark protocol.
• Existing canonical validation/test classification semantics remain intact.
• Evaluation IDs/configuration/provenance make protocol status evident without relying on free-text interpretation.
• Regression tests cover complete and incomplete human-benchmark classification cases alongside unchanged validation/test cases.
• Relevant evaluation/human-benchmark documentation describes the new protocol designation and stops directing current work through superseded issue Superseded: Establish the current Phase 1 human benchmark gate #3.

Constraints / Non-Goals

This issue is about provenance and protocol classification only. It does not collect human attempts (#12), run the active reference (#13), define the statistical pass/fail gate (#15), or alter validation/test statistical protocols.

Validation

Using synthetic/temporary fixtures or a non-evidence-consuming classification test path, verify that: (1) the exact full human-benchmark protocol receives the new canonical human-benchmark designation; (2) one missing map or a reduced policy/protocol configuration is rejected or marked noncanonical; and (3) existing full validation/test runs retain their current classification.