feat: add kanban PR review polling loop

[tw3akercc]

Summary

• Re-scopes the Kanban PR follow-up loop onto upstream's existing single review status instead of adding in_review / code_review / merge_ready.
• Moves PR-bearing task completions into review while preserving run metadata, avoiding completed_at, and preventing dependent tasks from unblocking prematurely.
• Adds hermes kanban pr-review-poll <task_id> plus parser/poller logic for gh pr view, gh pr checks, requested changes, unresolved review threads, closed-unmerged PRs, and seen-item dedupe.
• Updates CLI handling, docs, bundled skills, and tests for the simplified review-loop metadata shape.

Test plan

• python -m pytest tests/hermes_cli/test_kanban_pr_review.py tests/hermes_cli/test_kanban_cli.py tests/hermes_cli/test_kanban_db.py tests/hermes_cli/test_kanban_core_functionality.py tests/tools/test_kanban_tools.py -o 'addopts=' -q -> 462 passed
• git diff --check -> passed
• python -m compileall -q hermes_cli/kanban_pr_review.py hermes_cli/kanban.py hermes_cli/kanban_db.py -> passed

Notes

• This replaces the earlier multi-state PR (feat: add Kanban PR review lifecycle #19314) that was closed on design-scope grounds.
• The branch is pushed from fork tw3akercc/hermes-agent:feat/kanban-pr-review-poller-v2.

[teknium1]

Thanks for re-scoping this onto the existing review lifecycle. Two blocking issues need correction before this can provide the documented polling behavior.

Problems

• hermes_cli/kanban_pr_review.py:297 reads only latest_run() for PR metadata. Current main creates a newer, metadata-free review run in claim_review_task() (hermes_cli/kanban_db.py:3538-3559), and latest_run() selects that newest run (hermes_cli/kanban_db.py:8916-8923). After dispatch starts the review worker, pr-review-poll therefore cannot locate the original PR reference.
• hermes_cli/kanban_pr_review.py:332 requests conclusion from gh pr checks --json. The current CLI rejects that field; _gh_json(..., default=[]) suppresses the error and makes all checks appear absent.

Suggested changes

• Persist or search historical PR metadata across the implementation handoff and review run; add an end-to-end DB test for poll-after-review-claim.
• Use supported gh pr checks fields (bucket/state) and test the actual constructed command shape.

Automated hermes-sweeper review.

[teknium1]

latest_run() becomes the newly-created review-agent run as soon as the dispatcher claims the card, but that run has no copied PR metadata. The original implementation handoff is then older, so this raises before polling. Resolve the latest run that contains PR metadata, or persist the normalized PR reference across the review handoff.

[teknium1]

gh pr checks --json does not support conclusion (the CLI reports only bucket, completedAt, description, event, link, name, startedAt, state, and workflow). This command fails and the default [] path hides it, so check failures and pending checks are never detected. Classify the supported bucket/state fields instead.