[rusholme]

What and why

Task 5 of the unattended run housekeeping-2026-10-22 (specification): rapid_docs records what tasks 1 to 4 landed, as system behaviour with its numbers, and Ben's rulings of 2026-10-22 in decisions.md. Every earlier task's ledger line is TASK DONE, so everything is documented as landed; nothing is reported as not landed.

Pull requests documented (Caltech-IPAC/rapid, rebuild): task 1 bytes #312 (intake writes exposures only when expected), #313 (working directory removed after put), #314 (difference outputs trimmed and gzipped; packet reads either suffix), runbook #315; task 2 batch #316 (jobs.batch_job_id), #317 (the loop's reconcile, job submit --job), #318 ([batch.memory_mib]), runbook #319; task 3 pass #320 (listing by the selection's observations), #321 (one connection and one flush per publish pass), #322 (validate_run.py eight at a time), #323 (consumer check one set at a time), runbook #324; task 4 reference #325 ([reference.selection] before_mjd, held, the catalogs key's dataset), runbook #326 with "Results of part 8".

Files

• system/difference.md: the published files, each with the reader that earns it (the ruling's four: difference, uncertainty, mask, catalog; plus the three packet reads), gzipped FITS, the checksum of the gzipped file, no retention setting, a variant made again by a rerun, either-suffix reads, the cutout parity (781 of 781); SFFT and naive run for their bits only; the trimmed bytes in Measured; a "Not decided" bullet on whether SFFT/naive keep running.
• system/foundation.md: intro; the job row's Batch job id and the reconcile's reason; the catalogs key with its dataset in Part 1's key table; job submit --job in the verbs; the working directory's removal under "Where outputs live"; per-stage memory as a table under "A job on Batch"; the thin chain's intake, reference and difference rows; publish's one connection and flush, the consumer's one-set memory, the producer pointing at its ruling; the settings paragraph names [batch.memory_mib] and before_mjd; "On SMDC" has eight parts, the shared-field bullet points at its ruling, and a new "Housekeeping" part with five bullets; Validation gains the bounded script; Measurements gains "Housekeeping" (before and after, from the PR bodies and Results of part 8) with the disjoint reference's counts and three observations; "Not decided here": the reference-per-SCA, disjoint-reference and catalogs bullets leave (settled), the producer, other-filters and injection bullets are restated against their rulings, and enter: depth vs disjointness, the bound per filter, the psfs sequence against copied ids, SFFT/naive.
• system/loop.md: the reconcile before the snapshot, the listing by observation, the reference rule (per field and filter, one SCA's rows) and the difference rule under a bound, Kafka on one connection and one flush, job submit --job as a rerun, held and the reconcile in the pass line, a new "The reference bound" section, the per-field chain in admission order, the measurements before and after; "Not decided here" loses the four items that landed and gains three.
• system/stage-contract.md: the Batch job id and memory ask on submission, --job, refusal cleared at an attempt's start, the working directory removed after upload.
• system/intake.md: the conditional exposures row (the page said every intake writes it); its "Not decided" bullets on the reservation and the duplicate row updated.
• system/decisions.md: eleven rulings of 2026-10-22, one dated line each.
• system/index.md: scope cells for decisions, foundation, stage-contract, difference, intake, loop.

Divergences from the brief, with the reason

• intake.md is edited, though the brief does not name it: it stated that every intake writes the exposures row and listed that as undecided, both false after #312. The repository wins.
• One more ruling than the brief lists, decision-breadth-before-injection: Ben ruled on rapid-rusholme on 2026-10-22 at 23:03 UTC (16:03 PDT), after the brief was written, that the other filters run un-injected, PSF from each file's CRDS ePSF reference, before injection (recorded in the work log). Leaving it out would make decision-housekeeping-first the latest word on the order when it is not.
• Times on the rulings: PLAN.md v7 gives the stepwise session as 12:57 to 13:08 PDT with times only for the PSF (13:08) and Spot (13:23); the others carry the date alone.
• The reference per (dataset, field, filter) and the chain in admission order are one entry, as the brief lists them as one item.
• No run ids, job ids, Batch ids or bucket names on the site; the runbook (part 8) carries them. Instance types and the dataset name are used, as in foundation, intake, loop, stage-contract, decisions, index: the HLWAS run from SOC L2 files to the Kafka topic #91.

Build

Not built on the host (no Sphinx; provisioning is an install). docs.yml is the check. Locally: scripts/check-public-safety.sh OK; every {ref} decision target added resolves; the in-page anchor #the-reference-bound is a level-2 heading under myst_heading_anchors = 3; the added text carries no em or en dashes and none of AGENTS.md's excess-vocabulary words.

Measurements

Documentation only: no operation ran. The figures are tasks 1 to 4's PR bodies and the runbook's "Results of part 8".

Improvements seen (not applied)

• reference.md still states its scope as (field, filter, SCA) without saying the product key has no SCA and which SCA's rows the loop picks; the loop and foundation pages now say it, and a sentence there would close it.
• foundation.md's "The wide-area ramp" storage bullet says 42 uncompressed FITS per epoch while #314 counted 72 objects in a run 12 difference job; the bytes agree, the count does not. Left as the ramp's own record.
• specification.md, tool.md and operations.md still describe the processing-date loop (carried from foundation, intake, loop, stage-contract, decisions, index: the HLWAS run from SOC L2 files to the Kafka topic #91).

What the next task must know

• New ruling anchors: decision-housekeeping-first, decision-breadth-before-injection, decision-reference-bound, decision-chain-admission-order, decision-injection-outside, decision-producer-last, decision-catalogs-dataset, decision-difference-trimmed, decision-gbtds-after-injection, decision-psf-from-epsf, decision-spot-parked.
• Status stays DRAFT on every page.

Left

Nothing in this task's sentence.

🤖 Generated with Claude Code