fix(kanban): harden worker runtime ownership and preflight

[stevenbaert]

Summary

This hardens the Kanban worker lifecycle so stale process IDs, reused PIDs, gateway restarts, and inherited provider routes cannot corrupt worker outcomes or terminate an unrelated process.

What changes

• Bind worker exit evidence to a specific task, run, process identity, and process group.
• Require current in-memory ownership before timeout/reclaim termination; otherwise fail closed.
• Preserve bounded, redacted crash context for real process exits.
• Validate workspace, profile, provider, model, credential reference, toolsets, and skills before spawning.
• Persist the effective provider/model route and deterministically inherit it for reviewer/child tasks.
• Prevent stale exit evidence from being consumed by a newer run.
• Add regression coverage for PID reuse, gateway restart, ownership checks, route inheritance, invalid models/credentials, salvage, and POSIX/Windows termination paths.

Verification

Rebased onto current NousResearch/hermes-agent:main and run from a clean worktree:

scripts/run_tests.sh \
  tests/hermes_cli/test_kanban_worker_runtime_contract.py \
  tests/hermes_cli/test_kanban_db.py \
  tests/hermes_cli/test_kanban_core_functionality.py \
  tests/plugins/test_kanban_worker_runs.py -q

436 passed, 0 failed
python -m py_compile hermes_cli/kanban_db.py  # passed
git diff --check origin/main...HEAD          # passed

Independent exact-head QA also reproduced the two prior PID-reuse races and validated the canonical provider/model/auth and parent→reviewer inheritance paths.

Risk and scope

The changes are limited to Kanban worker lifecycle/preflight behavior and its tests. No gateway configuration, secrets, deployment, or runtime activation is included.

[teknium1]

Thanks for the detailed worker-ownership work. The current PR needs rework before its lifecycle changes can be salvaged.

Problems

• hermes_cli/kanban_db.py:6953-6978 restores immediate blocking for clean-exit protocol violations. Current main deliberately changed this to a bounded, violation-only retry streak in 452861fdc (hermes_cli/kanban_db.py:6821-6886), including per-task retry precedence.
• The ownership fail-closed rule is incomplete. The PR's _terminate_reclaimed_worker returns unverified ownership when its registry record is absent (hermes_cli/kanban_db.py:6313-6342), but release_stale_claims still releases the claim at :3715-3744; detect_stale_running has the same termination-attempt-only guard. An unverified live pre-restart worker can be overlapped by a respawn.

Suggested changes

• Carry the registry/process-group work onto current main's bounded protocol-violation path.
• Defer automatic stale reclaims when host-local ownership is indeterminate, and add restart-path coverage for TTL and heartbeat-stale reclaim.

Automated hermes-sweeper review.