Skip to content

docs: sweeping for the shape, and what a guard cannot ask (#2492) - #2506

Merged
gHashTag merged 1 commit into
masterfrom
docs/shape-sweep
Aug 23, 2026
Merged

gHashTag merged 1 commit into
masterfrom
docs/shape-sweep

Conversation

@gHashTag

Copy link
Copy Markdown
Collaborator

#2505 named a shape rather than an incident — a verification aimed one step to the left of what the change replaces. So every deploy path in reach got the same two questions: what does it REPLACE, and what does it verify?

Four paths, one hit.

The hit was the worst one available

publish-website.yml runs on a 15-minute cron, unattended, and replaces the site's assets/ with a build from another repository. It called the blog regenerator and the drift checker and nothing that asks whether the incoming build still carries the posts the site is serving — the exact gap that cost eleven posts two days of downtime, sitting in the one path that fires ninety-six times a day without a human present.

And the drift checker structurally cannot ask it

verify-site.sh compares the shipped bundle against the static tree and fails on disagreement. That is the right check. It is also blind to a slug leaving both sides at once, which is what happened: the posts lived only in the bundle, the bundle was replaced, and nothing disagreed with anything.

A checker of agreement between two artefacts says nothing about preservation across time. Different questions; one instrument cannot hold both.

The clean paths were clean for a copyable reason

The CNAME guard reads the CNAME file. The Pages guard reads the Pages API and curls production. Each verifies the artefact it names, and the second crosses an independent channel to do it.

Then the sweep turned on the instrument

The new gate counted slugs by PRESENCE on disk. rsync is additive, so yesterday's chunk sits beside today's: a slug dropped from the reachable chunk is still on disk in the unreachable one, so a disk-wide read calls it live while the site 404s it — and the orphan prune that follows a deploy then deletes it for good.

That is the 2026-08-21 mechanism exactly, and the instrument built to catch it shared its blind spot. Today the apex carries 16 chunks, one reachable, and it happens to hold all 48 slugs: the old reading was right by luck.

Reachability from index.html now. Eighth control case, verified RED under the old presence-based reading. The planted trees carry a real index.html and entry chunk so the cases run the walk rather than the no-index fallback — planting only the chunk would test the fallback and call it the gate.

Verified where it matters

In CI, not only locally: the cron fired on the commit that added the gate and printed no post is lost — live serves 48, this build carries 48, 0 new. And on real data the other way: the pre-restoration build exits 1 and names all eleven posts.

Refs #2492

@github-actions

Copy link
Copy Markdown
Contributor

📓 NotebookLM Notebook linked to this PR

This notebook contains session context, decisions, and artifacts for this work.

@github-actions

Copy link
Copy Markdown
Contributor

PR Dashboard

Generated at: 2026-08-23 06:32:12 UTC

Summary

Status Count
Total Open PRs 6
PRs with Failing Checks 2
PRs with All Checks Green 4
READY 3
FAILING 2
PENDING 0

Seal Status

  • ⚠️ STALE -- sha256(compiler.rs)=e905ba8eeb53 != manifest seal=87e5cbd3ad94.
    The committed NMSE numbers were certified against an older compiler.rs.
    Run scripts/reseal-check.sh locally for the two-step reseal command (advisory; not a merge gate).

@gHashTag
gHashTag merged commit 9d42d76 into master Aug 23, 2026
19 of 21 checks passed
@gHashTag
gHashTag deleted the docs/shape-sweep branch August 23, 2026 06:33
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant