Publish real LLM benchmark and 48/100 official patch resolutions

[YeDonS]

The project now has a completed real LLM experiment and official functional patch results for the same fixed 100 historical issues. This replaces the previous credential-blocked status with inspectable predictions, causal explanations, source citations, traces, and per-instance official reports.

• Real inference: 100 complete, 90 cited diagnoses, 10 abstentions; file Hit@1 87%, Hit@5 88%, MRR@5 0.875.
• Official SWE-bench 5.0.2: 48/100 resolved, 23 unresolved, 27 empty patches, 2 test timeouts (600 seconds), no missing outcomes. Four frozen batches preserve original task order and exact patch hashes.
• Automated post-freeze causal comparison: 87 correct, 1 partial, 2 incorrect, 10 unscorable. This is labeled automated agreement; human accuracy remains null.
• Added durable official reports, image digests, test excerpts, raw-log hashes, invalid candidates, failure analysis, real authenticated HTTP evidence, and updated English/Chinese README and reproduction documentation.

The actor used codex-cli:gpt-6.1-sol with medium reasoning and HTTPS through the CLI's existing local login. No authentication files or API credentials enter the repository or CI. Inference source stayed frozen at 3dc26ac14fe7573bb84ef851262645353842ef58; reference patches and hidden tests were only used after each prediction froze. Historical inference is source-only; separate real-model CLI/HTTP checks actually exercise all three roles on the trusted synthetic fixture.

Validation: 75 local tests passed; Ruff lint/format passed; sdist/wheel build and clean wheel installation reproduced the packaged fixture and found all 100 tasks. Verified all outcome IDs, frozen batch identities, patch/report/log hashes, source digest, JSON, documentation links, and credential-pattern scan. CI checks three Python versions and both Docker images. Official evaluation runs: batch 1, batch 2, batch 3, batch 4.

Limitations are retained: training contamination on public historical issues, same-model review bias, model alias drift, 17 invalid candidates, 10 abstentions, and 2 test timeouts. No outcome-based repair, exclusion, human accuracy claim, or controlled orchestration ablation is implied.

Two post-hoc maintainer-patch controls also passed official required regressions with the same image digests (control workflow). Incidental fixture/import errors still occurred elsewhere in their output, so those messages alone do not invalidate the grade. Both original model failures remain unresolved; controls are separate from 48/100.