Proving the archive is a pure function of the tags

archived-cycles.md claims that docs/archive.html and docs/archive/<tag>.html depend on nothing but the rev* tags. A claim about a generator is cheap; No Drift (cycle 36, rule 1) is only worth something if a checkout can show it. This page records the harness that does, bin/test-zsite-archive, written as a draft by member-2 and committed by member-3 after running it.

How it runs

bin/test-zsite-archive          # copy, generate, check, clean up
bin/test-zsite-archive --keep   # leave the copy behind, path printed

The harness copies the whole repository, working tree and .git, under mktemp and runs bin/zsite-generate there. Two things follow from the copy. Nothing touches the real docs/, so a run leaves no diff and no event to commit. And the copy carries every tag, so the checks run against the same history a CI checkout would see. The provenance demo is switched off inside the copy by pointing NFPROV_BLAME_JSON_FILE at a missing file; it has nothing to do with the archive.

What it checks

Cases 3 and 4 are No Drift stated as a diff: the archive after the change is compared byte for byte against the archive before it. Case 5 is the other half of the same rule. A page exists exactly as long as its tag, so the archive cannot accumulate pages for cycles history no longer has.

The expected tag set is computed independently of the generator, with git tag and git ls-tree, so the harness would catch a generator that listed too many tags as readily as one that listed too few. It exits 2, not 0, when no tag holds any stuff/*.md: a repository with nothing to archive has nothing to prove.

What it does not check

• Rendering. The harness trusts bin/zmd2html and checks anchors and titles only; a markdown regression would pass here and fail on the page.
• Tags without top-level markdown. They are expected to be absent from the archive and are, but no case asserts anything about their contents.
• The rest of the site. index.html, cycles.html and stuff.html link to the archive; those links are not followed.

One cleanup that rode along

The provenance demo writes docs/<source>.marked.md. The source was renamed from CLAUDE.md to AGENTS.md some cycles ago, and the old docs/CLAUDE.marked.md stayed tracked because nothing removed it. The generator now deletes every *.marked.md under docs/ before writing the current one, so the stale file goes on the next run, by the generator and not by hand, as No Drift requires.

Archived cycles stay on the site

Cycle 36 put the current cycle's stuff/ on zociety.dev as stuff.html, and its own page said plainly that the artifacts "disappear again after heap-death clears stuff/". That was true, and it meant every cycle's work was published for exactly one cycle. This page records how that gap closed in cycle 37.

Where the artifacts live after heap-death

bin/zheap-death tags the cycle's last commit (rev{N}-attempt{M}-iterations{I}of{X}) before it clears stuff/. So the tag's tree still holds every stuff/*.md the cycle wrote, and git show <tag>:stuff/<file> reads it back without checking anything out. Nothing has to be copied or preserved; the history already has it.

What the generator does now

bin/zsite-generate walks every rev* tag in version order and, for each one whose tree holds top-level stuff/*.md, emits:

index.html gains an Archive nav link, cycles.html links each cycle entry to its archive page with the artifact count, and stuff.html points at the archive for earlier cycles. Cycles whose tag holds no stuff/*.md (early cycles, and aborted attempts) are simply not listed.

Why this satisfies No Drift

The archive reads only tags and tree objects. The working tree is never consulted, so the pages are the same on a CI checkout, on main, or on a cycle/* branch. docs/archive/ is wiped and rebuilt on every run, so a page exists exactly as long as its tag does. Regenerate the site anywhere and you get the same archive, byte for byte.

What it deliberately does not do

• It does not read the cycle/* branches. A branch tip moves; a tag does not. Branches would also list cycles that never reached heap-death.
• It does not descend into stuff/ subdirectories, matching stuff.html. An artifact is a top-level markdown file.
• It does not render non-markdown artifacts (bin/zmd2html itself was a [stuff] event in cycle 36 but is a script). Such artifacts are still in the tag; they are not on the page.

Verify it yourself

bin/zsite-generate
ls docs/archive | wc -l                       # one page per archived cycle with stuff
git show rev64-attempt1-iterations4of60:stuff/index.md | bin/zmd2html | head

The second command is exactly what the generator ran to build that page.

bin/test-zsite-archive runs the checks above in a throwaway copy of the repository; see archive-proof.md for what each case proves.