feat(ops): restructure the weekly digest around three production windows

[modelmirror]

The weekly performance digest reorganised around what was produced, over three windows, with the boards left to the website.

What changed

The digest's blocks are now, in order:

1. Produced this week — unchanged: window bounds in the heading, cells and events and measured spend over one set of usage.json records, the role × stage table. The spend-backstop paragraph no longer closes it.
2. Produced this month — the same shape over the spend backstop's own trailing window (spend.window_days in config/tracking.yaml, 30 today), so the census and the verdict beside it are one window over one walk of the ledger; the backstop paragraph closes this block.
3. Produced this term — the same shape from 1 October of the Term the digest falls in (october_term_year), pinned to that midnight rather than counted back in days so the Term's first morning is never dropped; closes with Forward cells scored (<scope>, ledger to date): …, the frozen-process figure that used to be a health question, with the shakedown state still named when the frozen scope has nothing scored in either stratum.
4. Backtest results — unchanged, last.

The Health questions and Analytics state blocks are removed, with the helpers only they used. WeeklyAnalytics now carries only the three back-test artifacts and the scorer in force, so the digest no longer loads the leaderboard, claim-scores, big-cases or statpack files. cell_census and spend_over gained an optional since instant. The daily ops report (render_markdown) is byte-unchanged.

Docs: docs/pipeline.md (the block list and the vintage rule, now scoped to the back-test block), docs/cli.md, README.md, the ops-report docstring (which also named a flag that does not exist), and two comment lines in run-ops.yml (the header's block list, and a stale claim that the digest carried a stale-label signal; one comment also named a WEEKLY_TICK env that no step sets).

Example, rendered over the committed ledger

## Produced this week (2026-09-14 to 2026-09-21, the 7d before this digest)
**373** cell(s) over **128** event(s), and **$946.04** of measured model spend …
## Produced this month (2026-08-22 to 2026-09-21, the 30d before this digest)
**476** cell(s) over **145** event(s), and **$1,170.33** of measured model spend …
Spend backstop: **$1,170.33** of **$2,500.00** over the trailing 30d window (**47%** consumed, $1,329.67 left, 476 cell(s)) — clear …
## Produced this term (OT2025, since 2025-10-01 — 355d to 2026-09-21)
**1115** cell(s) over **223** event(s), and **$2,467.92** of measured model spend …
Forward cells scored (frozen, ledger to date): 0 total, no prior snapshot to diff. No frozen-process cells yet — still shakedown.
## Backtest results

Executed checks

• uv run fedcourts ops-report --digest-out digest.md --generated-at 2026-09-21T08:30:00+00:00 renders the above; the month block and the backstop line read the same 476 cells and $1,170.33 by construction, and a test now pins that equality on a ledger with rows inside and outside the window.
• The 1 October render is tested directly: OT2026, since 2026-10-01, a 0-day window that keeps the post-midnight cell and drops the 23:00 one from the night before.
• scripts/gate.sh lint, types, test (4970 passed, 2 pre-existing skips), schemas (no drift; no pydantic model moved), data — all pass. uv run fedcourts process-digest --all byte-identical to origin/staging.

Review

code-reviewer (no blockers; five recommendations, all taken: the ledger-to-date denominator on the rendered line, the docstring synced to the docs, the defusing test re-pointed at the back-test predictor ids since the census labels are enum-valued, the month-equals-backstop assertion, the 1 October parametrize row) and docs-reviewer (one blocker, README's block list, fixed; the ops-report docstring, the two workflow comments and _sourced's rationale also fixed). Not taken: making the window bounds required on ProducedWindow (the fallback is test-only; left for a later tidy).

Touches .github/workflows/run-ops.yml (comments only): waits for the maintainer, per the golden rules. Effect check for the promotion body: the next Monday digest (2026-09-28) opens with ## Produced this week as its first block and carries ## Produced this month and ## Produced this term (OT2025, …) before ## Backtest results, with no ## Health questions or ## Analytics state.

🤖 Generated with Claude Code