Add a complete agent-axis run (Fable 5 / Cursor agent × TensorCircuit-NG): 12/12 official reward 1.0, runtime-ratio protocol, and two infrastructure fixes

[QingyunQian]

What this PR does

1. A complete agent-axis benchmark run (results/fable5/)

Adds a full 12-task agent-axis data row: solver = Fable 5 via Cursor agent, framework fixed at TensorCircuit-NG. Every task passes the unmodified official verifier — functional evaluator, static policy, and Codex LLM audit (model gpt-5.6-sol, recorded per task) — with reward = 1.0 on all 12 tasks. Per-task artifacts include the solution, verifier outputs, audit provenance, and a same-machine runtime comparison against the publication reference. A full report (results/fable5/REPORT.md) documents the pipeline, per-task observations, and reproducible summary figures.

Harness caveat, stated upfront: the solver ran outside the Harbor agent container (Cursor agent operating on the workspace, verification through the documented verifier-only candidate-check path). Artifact-side metrics (pass components, runtime, T/T_ref) follow the standard verifier and are comparable; agent-side metrics (tokens, solve wall time) are not comparable with in-container rows. Solving-integrity protocol: only instruction.md and the framework prompt were read; tests/ and solution/ were never opened during solving; references were executed for timing only, after each candidate was frozen.

2. Fix: reward formula desync between docs and scorer

Since ea088db, AGENTS.md/README declare reward = functional × static_policy × llm_audit ("runtime … must not reduce pass reward"), but score_submission.py (template + all 12 task copies) still multiplied runtime_score into reward. Observed impact: a run with all three gating components at 1.0 and runtime_sec = 196.25 scored reward = 0.8645833… — exactly (300−196.25)/(300−180). This PR syncs the code with the declared policy (runtime_score/runtime_sec still computed and reported unchanged). Also available as a standalone atomic PR if preferred.

3. Fix: framework pin predates the omeco integration

frameworks/tensorcircuit/requirements.txt pinned tensorcircuit-nightly==1.7.0.dev20260618, which predates set_contractor("omeco") landing in master, so the publication references for tasks 01/05 crash on the public image (the omeco package itself was already listed). Bumped to 1.8.0.dev20260726: all 12 references now run unmodified; the four candidates already measured were re-verified on the new pin (all still functional 1.0).

What we found

• Artifact efficiency (T/T_ref, same machine & image): official-record median 2.39. At matched precision the candidate beats the expert reference on two tasks (06: 0.76, 11: 0.74), is within 1.3× on two more (05/07), sits at 1.7–3.6× on six (mostly XLA-compile/contraction-path overhead), and shows two large gaps (04: 11.9, 10: 18.1 — the 18-qubit CZ hyperedge via multicontrol-MPO is ~18× slower than the expert treatment).
• Precision attribution: all 12 references use complex64 (checked post-freeze, configuration lines only). Five candidates ran complex128; their complex64 controls were re-run through the official evaluator. Notably, task 04's control fails functionally at single precision (the 2e-4 channel-fit tolerance is unreachable with the vectorized-DM approach) while the reference passes at c64 — so that ratio partly encodes an algorithmic precision-requirement gap, not just speed.
• LLM-audit boundary calibration: task 04's first implementation (a numerically-validated hand-built Kraus-ladder network on raw tensornetwork nodes) passed functional+static but was rejected by the Codex audit as a raw-simulator bypass; the framework-native rewrite on tc.MPSCircuit (numerically identical to 3e-15) passed and runs 4× slower — a concrete quantification of the "compliance tax" the benchmark is designed to surface.
• Framework observations collected in the report: default-contractor pitfalls (full-wavefunction materialization at 49 qubits; MPO-sandwich hangs), JIT constant-capture of a 2.9 GB sparse Pauli-sum at 22 qubits, cond_measure/QuditCircuit/multicontrol working well as framework-native primitives, and a measurement-insensitive protocol emerging from trajectory-averaged feedback optimization (task 07).

Note for maintainers

• The default audit model in AGENTS.md/README/conf.toml was updated from gpt-5 to gpt-5.6-sol during the run (the ChatGPT-auth catalog no longer serves the bare gpt-5 slug). Keep or revert as you prefer.
• results/fable5/ placement and naming are proposals; happy to move/rename if you want a different results layout.

[refraction-ray]

this figure has no extra info than the other one?

[QingyunQian]

As it is Cursor agent ,which means different harness for the task, so transfer to draft and dont merge now