feat: add operator-reviewed infrastructure replacement runs

[evalstate]

Summary

• Add operator-reviewed infrastructure replacement runs: select exact native errored trial IDs, enter/review a budget, and submit through normal authorization, inference-binding, idempotency and lifecycle protections.
• Preserve original execution and artifacts. Native repeated task configurations execute the selected multiplicity, including one-of-many repeat attempts. Disjoint replacement batches and replacement-of-replacement review are supported without score-based selection.
• Use public Harbor JobPlan.aggregate for the combined native result. The run page separates Original, Replacements and Combined, with explicit selected-cohort versus all-incurred reported cost coverage.
• Exclude related subsets from independent leaderboard pooling; pending/invalid assemblies fail closed. Only unambiguous native scalar means populate the existing leaderboard field.
• Includes merged PR fix: display Workbench providers and bump fast-agent starter to 0.10.24 #219 through the main baseline: recorded Workbench provider display and the fast-agent 0.10.24 starter remain intact.

Native ownership and approved bridge

Checked Harbor dcd0a7ac74b7bd417780d9cb27cd819c7ec82e4e and relevant history, including the public JobPlan extraction and regrade derivation:

• src/harbor/job_plan.py: resolve_task_configs, build_trial_configs, resolve_metrics, public constructor, aggregate, aggregate_stats.
• src/harbor/job.py: Job.create, Job.run, native expansion/resume and regrade construction.
• src/harbor/models/job/{config,lock,result}.py and models/trial/{config,result}.py: native configuration, identities, locks, results and accounting.
• src/harbor/tasks/client.py: repeated task download ordering.
• src/harbor/metrics/{base,mean}.py and utils/pass_at_k.py: native metric shapes and calculations.
• src/harbor/hub/client.py: hosted relaunch/combined behavior is not the existing Bucket integration.

The only new durable fact is optional RunRecord.operator_selection: an operator review identifying an existing run, exact native trial IDs and a native source-evidence fingerprint. This is the separately approved, revision-scoped bridge for cross-run selection/provenance/coverage, not a renamed native regrade action. Replace it at the first reviewed Harbor revision providing the equivalent storage-neutral contract; no removal revision has been identified.

No new scheduler, metric implementation, trial identity, result format, SQLite table, persistent resource, source-artifact rewrite or execution-state override. Parent preflight independently verifies the same trusted ancestry before Job.create. Existing compiled components are inherited exactly; direct-config catalog restrictions are not broadened.

Validation

Author suites passed: 1,523 TypeScript tests, 75 browser tests, 394 agent-package tests, and 102 CLI tests. New backend modules and the UI exceed 85% coverage; changed Python modules exceed 85%. Independent review found two high-severity issues, both fixed and re-reviewed:

• Native sha256: source fingerprints now agree across Python, TypeScript transport, durable schema, ancestry and parent preflight. A real cross-language regression covers that path.
• Ambiguous native multi-reward/multi-metric objects are withheld rather than incorrectly treated as a scalar mean.

Final feature-snapshot validation is complete: root coverage 89.10%; Ruff/format/ty, formatting/lint/typecheck/build, dependency audits and committed generated-output checks passed. Both normal-network linux/amd64 Docker builds passed locally and in feature-head CI. Ordinary Slophammer scanned 27/27 production files without findings; DRY reported zero candidates. Final independent design review: GO. The documentation-only final evidence addendum also passed fresh complete CI at exact head 71bde2cd4dcd759f010499171804921c4cb32ff3, including both Docker builds.

Limits and operations

• The existing package-wide agent coverage gap remains; no coverage threshold is weakened. The repository still lacks a Slophammer baseline file; normal check/DRY checks are used in addition to recording that failed required baseline invocation.
• Custom or ambiguous native metrics remain available in the full combined view but cannot populate the current scalar leaderboard without a separate representation decision.
• Uncertain submission retries retain their key/body across panel closure, not page reload; immutable overlap protections remain server-side.
• No paid replacement execution, live private-source canary, credential movement, or benchmark control action was performed. Inspector/source transport is mocked in offline cross-language tests.
• Implementation, merge, matched image publication and deployment were separately authorized. Actual replacement execution still requires operator selection and budget approval. Do not interrupt existing benchmark or setup Jobs during cutover.