fix: run parents locally and consolidate dashboard refresh feedback

[evalstate]

Summary

• Run the Harbor parent on ordinary local /data storage; remove its Bucket volume and use acknowledged Hub SDK copies to the existing artifact Bucket.
• Restore native files before Job.create()/native resume. Keep control-owned records read-only to the parent; refresh control intent through the API.
• Preserve dashboard observations with native job metadata, trial config/lock identity, completed-trial results, and attempt-cost receipts. No new result schema, progress engine, scheduler, or resource.
• Invoke Harbor's existing scrubber on the actual trial before END callbacks using the separately approved, revision-pinned private integration. Retain native final scrubbing; do not serialize unsanitized hook results.
• Publish each copied subtree's result last. Use Harbor's native completion lock for aggregate snapshots/cost checks. Withhold failed retry-enabled terminal results until final publication so interrupted native backoff cannot restore an obsolete completed attempt.
• Exclude the global job.log, which is outside native trial-directory sanitization. Preserve scrubbed per-trial logs.

Additional committed work

• Includes all six additional local commits requested for this PR: the stable,
  consolidated Run detail refresh strip, cancellation-test cleanup, and their
  authorization history.
• The strip observes existing query-cache state without adding polling or
  reinterpreting native Harbor evidence. Error details and retry access remain
  available in a collapsed overlay without page movement.
• Concurrent uncommitted provider-display and starter-version implementations
  are not part of this PR; their approval records are included without silently
  publishing the in-progress edits.

Harbor boundary review

Pinned dcd0a7ac74b7bd417780d9cb27cd819c7ec82e4e:

• src/harbor/job.py: Job.create, Job.run, on_trial_started, on_trial_ended, native resume and _trial_completion_lock.
• src/harbor/trial/trial.py, trial/hooks.py, trial/queue.py: actual Trial lifecycle, _emit, _scrub_jobs_dir, TrialHookEvent, and retry cleanup/backoff.
• src/harbor/models/job/config.py, models/job/result.py, models/trial/result.py, models/job/plugin.py: authoritative configuration/results and plugin contract.
• src/harbor/verifier/verifier.py, environments/hf_sandbox.py, environments/tar_transfer.py: awaited transfer followed by local reward lookup.
• src/harbor/cli/plugins/harbor_hub.py, cli/job_plugins.py, upload/uploader.py, upload/storage.py, storage/base.py: native hosted upload is tied to Harbor's service; it is not an HF Bucket transport for the existing file layout.

Relevant pinned history was reviewed; the trial source at e1be9bd39c368f88a27fee4cbd26655e9994ee5a still dispatches END before scrubbing. Replace this private integration when a reviewed pin supplies a supported post-sanitization hook or equivalent native ordering. Revision mismatches fail explicitly. Native configurations, identities, costs, rewards, and resume remain authoritative; durable schemas are unchanged.

For the console change, also checked native JobResult, JobStats,
update_progress_counts(), TrialResult, and TimingInfo in the model files
above, plus src/harbor/viewer/scanner.py (get_job_result, get_trial_result,
configuration readers) and viewer/server.py job/trial GET endpoints. The new
UI represents browser fetch feedback only; no schema, API, persisted value, or
Harbor progress meaning changes.

Validation

Combined GitHub CI passed at 0ec09b0c2d94fdfe3d545dc38d3c22487cfb6f87.
Combined local validation excludes concurrent uncommitted workspace edits:

• Root Ruff, formatting, ty; 102 tests, 89.10% coverage; pip-audit.
• Agent Ruff, formatting, ty; 321 tests; changed runtime modules 91.44% coverage.
• TypeScript formatting/lint/typecheck, 1,458 tests, build, generated-contract check, npm audit.
• 73 Playwright tests.
• Both combined Dockerfiles built for linux/amd64; offline Python 3.12 native-binding smoke passed.
• Normal Slophammer check and DRY check; public privacy and full diff review.

Disclosed validation limitations

• Package-wide agent coverage is 71.84%, below the requested 85% target; the existing broader coverage gap remains. No threshold was weakened.
• slophammer-py check . --baseline cannot run successfully because the repository has no slophammer-baseline.json. The ordinary check reports no findings; DRY reports zero candidates.
• No paid Job, remote canary, deployment, credential movement, or upstream publication performed.

Operational trade-offs

Transfers are acknowledged snapshots, not a cross-object transaction or streaming log service. Abrupt termination can lose unacknowledged outputs/cost evidence. Errored retry-enabled trial diagnostics are conservatively delayed until final publication, and incomplete snapshots can repeat work under native resume. Native text scrubbing retains its existing best-effort limits, including skipped binary/unreadable files.

Deploy a matching reviewed worker image and control launch template together; mixed versions fail early. No compatibility mount or dual writer is added. Merge and matched parent-image publication/control deployment were separately
authorized. Cutover remains gated on idle execution; no running benchmark may
be interrupted and no benchmark launch or credential movement is authorized.