Standardize the Self-selected Top50 benchmark

[isPANN]

Closes #65
Closes #66
Closes #67
Closes #68
Closes #73

Summary

This single PR replaces the previous four-PR stack and delivers the complete Standardized Model API / Self-selected Top50 benchmark:

• introduce the problem-reductions repository and round-trip bug semantics directly in the triage prompt, then freeze a source-only Top50 with concrete hypotheses
• run 50 fresh sequential isolated rule episodes with fixed logical limits and exactly two submit attempts per selected rule
• define the benchmark limits once in code; remove the external budget file and redundant contract/policy version identifiers
• keep configuration claims out of submissions while validating actual private ledgers against the built-in limits
• package every shell and pred result with deterministic ANSI/progress/repetition cleanup, diagnostic-first previews capped at 10,000 characters, and read-only bounded raw logs
• bind exact observation metadata bidirectionally to shell and pred actions and reject outcome, command, size, or record tampering
• independently re-score accepted certificates and publish aggregate-only results
• isolate historical whole-repository results from the standardized Top50 table using the existing agent mode
• expose no logical-budget configuration and reject custom prompt, strategy, backend, model-kwargs, or pin overrides before scored model access
• verify baked source bytes, target marker, and trusted pred path
• retain the internal non-ranking calibration record and offline consistency checker
• migrate runner, schemas, CI, docs, skills, Docker metadata, and leaderboard presentation

Calibration scope

The checked-in grid is a human-reviewed bounded-prefix replay record from privately retained internal pilot trajectories. The offline command verifies the checked-in record, complete grid, selected knee, report, and built-in parameters; it deliberately does not claim to reconstruct private raw trajectories in CI.

Verification

• pytest -q — 333 passed, 5 skipped
• python3 -m benchmark.calibrate_budget --check benchmark/docs/budget-calibration.json
• changed-file Ruff checks
• git diff --check
• deterministic common-command demonstration recorded under runscribe/A-observation-packaging/A1-policy-shortens-noisy-common-commands/runs/2026-07-20T17-24-32__final-policy-clean

Do not merge until CI is green and the user explicitly approves.