[rusholme]

docs/foundation-smdc-step.md becomes a state page: what each of the thirteen parts proved, the current procedure, and the run log left to git history (git show 940e29b5:docs/foundation-smdc-step.md, git log -p). mig-runbook task 1 (planner brief, queue line mig-runbook).

What and why

The page was a 9,076-line run log at 940e29b5; it is now 554 lines:

• A short opening: what the SMDC step is, where the history is, the attribution rule (every number names its run and conditions; a superseded number points to the later part, which alone states the current value), and one sentence on where a new step's record goes (one row in part 13's table, plus in-place edits to Current procedure and What remains).
• ## Part 1 to ## Part 13, one section each: what the part proved, its numbers, the PRs. Step numbers the code and rapid_docs cite (part 7 step 6, part 9 step 2, part 10 step 5, part 11 steps 1, 2, 7, part 12 steps 1, 3, part 13's named steps) stay findable.
• ## Current procedure: the environment, bring-up of a fresh public with its grants and check, the image, a run (pointing to rapid_docs' loop page, Starting a night, for the verbatim recipe), the reference bounds of every dataset, the check, the gate, the exit codes, same-bytes comparisons and the instance-type caveat for packet equality, never prove on :85 or :86, the settings in force. Every command checked against the tip's --help (run, job local|submit|run|wait, official change|undo, publish, loop).
• ## Validation (made current: validate_run.py's checks at the tip, the catalog check in place of the chain, part 6's outcomes attributed to runs 9 and 10) and ## If a step fails (made current: job submit --job, CloudWatch INFO lines, rp job local to reproduce a difference).
• ## What remains: part 13's list, plus the earlier parts' items no later part closed, each named with its source.

scripts/check-public-safety.sh: the allowlist entry keyed on this file (634500218910, a Batch job UUID fragment in part 6's old table) is gone with the table, so the entry is removed, and with it the script's own self-entry for the same string, which no file now holds (the script's comment: an entry allows one file's matching string; anything new blocks).

Files

• docs/foundation-smdc-step.md
• scripts/check-public-safety.sh

Per part: kept, and superseded numbers dropped

part	kept	superseded or dropped
1	the replace (108 objects cascaded, 1 s, ops untouched), the grants (now Current procedure's bring-up, with part 6's rapid_read lines), the prototype ruling	the pg helper and the before/after row counts; the earlier draft's grant-count method
2	28 objects, 1,363,479,122 bytes, 34 s; write-once bucket	the upload loop (history)
3	run 1's chain, row counts equal to PRs 279 to 281, reader CLEAN; the default-privileges pointer (database/2026-10-24-object-catalog.sql cites part 3 for it)	the id plan, scopes and per-job tables; the associate stage marked retired (part 11 step 7)
4	run 4 on Batch, change group 9 / undo 10, 82 and 73 packets CLEAN, 1,348 / 1,324 s on one vCPU (attributed: that field, 1 vCPU, 15 GiB)	the per-job Batch ids and timings; the reference coadding its own epochs, pointed to part 8 step 4; the per-stage cost pointed to part 8
5	run 5's uncapped counts, the 25 min chain on one vCPU, the cap not a cost, change group timings; the determinism finding (runs 6, 7) and the seed	the member-chain depth table (chain retired, part 11 step 7); rubrat / NED (NED code deleted, dead-trim); the hour-per-image remaining-measurements list
6	runs 9 and 10, the PSF kind, seeded equality across instances, dev comparison (correlation 0.961 to 0.978, MAD), checks 1 to 10, group A 145 keys 256.9 ms, database 239 MB, run 10 200.7 GB	per-epoch dev table, per-group timings, per-run S3 sizes; run 10's bytes pointed to part 8
7	run 12: 426 epochs, 70 fields, 55,436 packets; steps 1, 6, 7, 10, 11	the open items now closed: bytes per epoch (part 8), dev's copied PSF residuals (parts 9, 10), the two-SCA field (ref-overlap), the chain (part 11 step 7)
8	the before/after table, trimmed to the rows that still stand	the consumer check's 2,397 → 2,740 s (part 11 step 5's 4 s stands); packet medians; per-step S/N residual counts (part 10 supersedes); the open list (gzip read cost carried to What remains)
9	the psf stage, 0.23 → 0.16, breadth totals; step 2's bound table (cited by rapid_docs' loop page and COMMON), F146's bound added as a row	the per-filter S/N > 20 table (part 10 supersedes); the reference-PSF rule (part 10 step 3); the CRDS 504s (part 11 step 4's fetch); open items carried or closed
10	steps 1 to 5 by name, the before → after table, the healthy ranges (cited by [gate]'s comment)	the step commands; star_sigma_px as first stored (PR 352)
11	one row per step, the gate's three bounds	alerted 0.40 kept but pointed to part 12 step 1; idempotence off, Spot capture, sequences (closed part 12); SCAs 7 and 9 (part 12, finder-mask); the night's session notes (history)
12	one row per step, the grants closed	SCA 7/9's diagnosis pointed to finder-mask; session notes (history)
13	one row per step: what stands, runs, PRs; superseded rows say so (differencer → diff-flavor; differencer-measure and dev-prefilter → sfft-rearm for HLWAS; flight-first's hand reference → ref-overlap; gbtds-second's 1,278 → gate-sci-psf's 1,196)	the "before" column; the live-run name list and job-definition list (COMMON and the plan hold them); session notes (history)

Divergences

• The brief's order (parts, then procedure) is kept; Current procedure follows part 13 rather than opening the page.
• database/2026-10-24-object-catalog.sql cites "part 3" for the default privileges, which were part 1 step 3; part 3's section names them, so the reference resolves and the dated file is not edited (implementation choice: no change to an applied schema file).
• Code and settings comments cite "runbook part N step M" (catalog.py, memory.py, metrics.py, intake_openuniverse.py, the overlays, foundation-thin.toml's "Results of part 10"); each cited step is still named in its part's section, so none was edited (out of scope: code and settings).

Measurements

	before (940e29b5)	after
lines	9,076	554
## Part N sections	13 (plus 12 results sections)	13
scripts/check-public-safety.sh	clean	clean

References at the tip, each resolving: database/2026-10-24-object-catalog.sql:12 (part 3), database/README.md:22 (the dated files' rule, the file), scripts/validate_run.py:7 (## Validation), scripts/check-public-safety.sh (entry removed with its string); rapid_docs system/foundation.md:844, system/loop.md:70 (part 9 step 2), system/datasets.md (part 12 step 3, part 13 steps): the rapid_docs PR fixes the sentences that described the page as a procedure of eleven parts and as holding full tables.

Improvements seen

• The loop page's exit-code table says "as on rebuild at efc03fc6" (rapid_docs system/loop.md); a later behaviour change will need the stamp moved; seen, not applied.

Session notes

• The brief's reference list missed rapid_docs' other citations of the runbook (loop.md:70, datasets.md:140,178,212, foundation.md:1207,1474,1510); a git grep -i runbook over both repos found them.

What the next task must know

A new planner step's record is one row in part 13's table (what stands, runs, PRs), and in-place edits to Current procedure and What remains where the step changes them; no step sections, no results section.

Left

Caltech-IPAC/rapid_docs#132 (four sentences of system/foundation.md). The plan's Operations facts line is the planner's.

🤖 Generated with Claude Code