feat(benchmark): add EdgeBench adapter contracts

[steven-kid]

Summary

• add a fail-closed EdgeBench/SForge readiness contract for one public task
• add a bounded no-execution run plan and compact final_result.json reducer
• expose edgebench-preflight, edgebench-run-plan, and edgebench-result-reduce
• document SForge ownership, 12-hour/cost gates, and the public evidence boundary

Validation

• python3 examples/edgebench-benchmark-adapter-smoke.py
• python3 examples/benchmark-core-adapter-contract-smoke.py
• python3 examples/cli-benchmark-dispatch-command-modularization-smoke.py
• python3 examples/benchmark-developer-workflow-doc-smoke.py
• python3 examples/cli-agents-last-exam-command-modularization-smoke.py
• PYTHONPATH=$PWD workspace/.venv/bin/python -m pytest -q tests/benchmarks tests/test_benchmark_ledger_countability.py tests/test_cli_entrypoint.py (62 passed)
• ruff check and ruff format --check on changed Python surfaces
• loopx check public boundary scan: 6 files clean
• CLI output base/head differential on a complete exact-main archive: 97/97 rows, no candidate-only rows, no review signals

Holds

• no EdgeBench task, Docker container, model API, upload, or submission was run
• standard premerge selected 18 checks; 17 passed and the output-budget canary timed out only while materializing an origin/main worktree from the local partial clone. The same differential passed in 23 seconds against a complete archive of the exact base commit.
• benchmark-sensitive changes require maintainer review; this PR is not eligible for self-merge

[huangruiteng]

Request changes

Reviewed exact head 8fd614ae. The no-execution boundary and CLI wiring are well done, but the compact result reducer can mark an impossible metric as countable.

Findings

• P1: reduce_edgebench_final_result accepts any finite best_pass_rate, including values outside [0, 1]. loopx/benchmark_adapters/edgebench.py:230-231 reads it with _optional_number, and the blocker block at lines 244-256 never checks range. I reproduced best_pass_rate=1.5 with valid round counts and runtime; the function returns countable=True with first_blocker="compact_edgebench_result_ready". A pass rate above 1 is not valid EdgeBench evidence and must fail closed before the compact result can be counted.

Minimum fix: reject best_pass_rate outside [0, 1] (and negative values) with a stable blocker such as edgebench_pass_rate_out_of_range; add negative smoke coverage for 1.5 and -0.1.

Validation

• examples/edgebench-benchmark-adapter-smoke.py: passed.
• tests/benchmarks + tests/test_benchmark_ledger_countability.py + tests/test_cli_entrypoint.py: 62 passed.
• benchmark-core-adapter-contract-smoke.py, cli-benchmark-dispatch-command-modularization-smoke.py, benchmark-developer-workflow-doc-smoke.py: passed.
• ruff check, py_compile, git diff --check: passed.
• GitHub pytest/build on this head: passed.

Merge decision: hold until the pass-rate range validation and negative smoke are added.

[steven-kid]

Addressed the requested pass-rate validation in 6e85bcc4.

• reduce_edgebench_final_result now adds the stable blocker edgebench_pass_rate_out_of_range whenever best_pass_rate is outside [0, 1].
• Added focused negative coverage for both 1.5 and -0.1.
• Boundary values remain valid: 0.0 and 1.0 stay countable when the other compact fields are valid.

Validation:

• python3 examples/edgebench-benchmark-adapter-smoke.py
• direct reproduction: 1.5/-0.1 -> countable=False, 0.0/1.0 -> countable=True
• python3 examples/benchmark-core-adapter-contract-smoke.py
• python3 examples/cli-benchmark-dispatch-command-modularization-smoke.py
• python3 examples/benchmark-developer-workflow-doc-smoke.py
• focused pytest: 62 passed
• Ruff lint/format, py_compile, git diff --check, and public boundary scan passed
• standard premerge automatic checks: 18/18 passed; benchmark-sensitive maintainer review hold remains

[huangruiteng]

Approved

Reviewed exact head 6e85bcc4. No blocking findings.

The prior pass-rate validation gap is fixed: best_pass_rate outside [0, 1] now returns countable=False with edgebench_pass_rate_out_of_range, and boundary values 0.0/1.0 remain countable. Negative coverage for 1.5 and -0.1 is included.

Validation

• examples/edgebench-benchmark-adapter-smoke.py: passed.
• examples/benchmark-core-adapter-contract-smoke.py, cli-benchmark-dispatch-command-modularization-smoke.py, benchmark-developer-workflow-doc-smoke.py: passed.
• tests/benchmarks + tests/test_benchmark_ledger_countability.py + tests/test_cli_entrypoint.py: 62 passed.
• Direct reproduction: 1.5/-0.1 -> countable=False; 0.0/1.0 -> countable=True.
• ruff check, git diff --check: passed.
• GitHub build/dependency-review on this head: pass; pytest pending at review time.

Merge decision: approved after reviewer.