fix: stabilize replacement evidence and contain parent errors

[evalstate]

Summary

• Normalize exact integral JSON floats before native evidence validation so actual JavaScript request transport preserves existing replacement fingerprints. Source artifacts and stored fingerprints remain unchanged; meaningful changes still fail closed.
• Pause on any unacknowledged owned HF parent ERROR instead of relaunching it. Existing explicit resume acknowledges observed errors under the run lock; new errors pause again. Native completion, cost-stop priority, and live-parent handling remain authoritative.
• Add cross-language regression coverage, 97 meaningful offline boundary cases, and an official reproducible empty Slophammer baseline.

Harbor-first review

Pinned Harbor: dcd0a7ac74b7bd417780d9cb27cd819c7ec82e4e (unchanged).
Checked native src/harbor/job.py, job_plan.py, models/job/config.py, models/job/lock.py, models/job/result.py, models/trial/result.py, models/verifier/result.py, and environments/hf_sandbox.py, plus their pinned history including native preflight/config validation and HF Sandbox provider/command execution changes.
Public APIs checked: Job.create(), Job.run(), JobPlan.from_resolved, build_trial_configs, aggregate, RetryConfig, and VerifierResult.rewards (float | int).
Harbor owns native validation, planning, trial retries, execution and results. Input normalization precedes native validation so Harbor restores declared numeric types. The sole optional durable field, acknowledged_parent_failures, records an HF-owned operator acknowledgment, not mirrored provider status or Harbor retry state. No new API endpoint, projection table, resource, native field alias, alternate hash or result writer.
See docs/replacement-evidence-parent-containment.md for ownership, checked history, compatibility limits and bridge removal condition.

Validation and independent review

• Independent production-code GO: 238 focused tests; no HIGH/CRITICAL code blocker.
• Independent incremental GO: 598 agent tests, 87.33% coverage; all 97 additions reviewed; empty baseline reproduced; zero findings across 27/27 files and zero DRY candidates.
• Final author gates: root 102 tests / 89.10%; Node 1,575 tests; 76 browser tests; Ruff/format/ty, Node formatting/lint/types/build/generated checks, dependency audits and public privacy checks pass. Both Dockerfiles built for linux/amd64. Node lint passes with existing warnings.
• Earlier whole-agent coverage and missing-baseline failures are retained truthfully in documentation and are resolved, not waived. No exclusions or thresholds weakened.

Release boundary

Release is exact-green-head squash merge, matched immutable worker image publication, and deployment to the existing private control Space with writes temporarily disabled. Fresh idle, provenance and unchanged execution checks gate cutover.
No benchmark launch/resume, replacement submission, inference grant, credential movement/change, paid canary or new resource is included. Deployment does not resume paused work. Offline preflight evidence is not a live-trial claim.