chore(grading): plan and authorize the regrade of the published selection

[kamui]

Why

The selection is settled (#132), the audit is declared (#128) and the queue's tools are in place (#129, #130). The owner authorized the regrade on 2026-10-12 after the grader-model test (#133): Claude Opus 5.5 at high effort, all 198 batches, a ceiling of $500 at list price. This records the plan and the authorization the queues run under.

Scope

• docs/research/cohort-rebuild-2026-10-05/regrade/approval.v1.json: the owner's answer, with the question and the other options.
• source-plan.*.v1.json, execution-plan.*.v1.json, authorization.*.v1.json for the queues selected (45 batches, ceiling $117.40), rebuilt (125, $312.87) and plain (28, $69.71), written by plan.py --cap 500.
• README.md there: what is graded, the estimate (about $337 expected, $603.30 of allowances), and how the owner starts the pilot and the queues.

Tradeoffs

• The ceiling is under the sum of allowances. An allowance is released down to actual cost when its batch finishes, and a queue reserves a batch only when it fits.
• Each queue's ceiling is its share of the $500 by allowances, so one queue cannot use another's room. If a queue reaches its ceiling it stops as budget-stopped and needs a new authorization.
• The grades are kept on a working branch until the audit's sample is drawn, so nothing here puts a grade on main.

Blast Radius

Research files only. The authorizations pin 23 runner files at this commit; a change to any of them before or during the regrade needs a new authorization version.

Verification

• plan.py without --cap and with it print the same queues: 198 batches, 663 saved reviews, 2,153 comments, allowances $603.30.
• Every pinned runner hash in the three authorizations equals the file in the tree (run).
• bun run verify:import and current_grading.py check pass; 0 of 623 admitted reviews assessed.
• The offline preflight passed for every batch earlier today with each queue's cache replacements (189 before feat(grading): take the review-code reviews into the published selection and drop three setups #132, and the 11 batches feat(grading): take the review-code reviews into the published selection and drop three setups #132 changed).
• No model was called.

Part of #30

[kamui]

Approved (advisory) — no findings.

Reviewed c3fd4d8. All 11 changed files reviewed, including exact regeneration of every source plan and execution order, input pins, budget shares, and pilot/resume behavior. No actionable findings or open questions.

Review details

Intent: Record the approved Opus 5.5 high regrade of the published selection under a $500 list-price ceiling.

Issue fit: Matches the planning and authorization portion of #30. Dispatch, grading and the declared audit remain later work.

Reviewed: c3fd4d8 against merge-base aa6ff1e.

Checks:

• Passed local exact-head plan regeneration and pin/budget checks: 198 batches, 663 saved reviews, 2,153 comments; combined queue ceiling $499.98.
• Exact-head CI passed preserved-evidence, current-grading and claim validation. CI run.
• Independent verification challenged cohort integrity, authorization accounting, and pilot/resume behavior.
• Reviewed as the authenticated PR author because the base has no reviewing-app declaration. Published as an advisory COMMENT.