Publish a retrievals ledger comparing consecutive pins as paired differences

[ChelseaKR]

What

There has been one retrieval (2026-08-07); the second will replace it, and the pages will describe the new file as if the old one never existed. This keeps every pinned retrieval's artifact under site/data/history/<source>/<retrieved>.json, adds site/history.html, and for each consecutive pair publishes the paired per-value differences a reader needs: records per year added or removed (FRAP), IRWIN coverage per year, per-field three-state counts, DINS incidents added, and per-incident completeness for incidents present in both. Each row carries both retrieval dates, both values, and the difference; no line is fitted, no direction is stated, and a field absent from one pin is shown as absent.

Why it matters

FRAP publishes versioned releases and the DINS layer is appended to after each season, so "what did the publisher fill in since last year" is the question a downstream user asks and cannot answer from either file alone. The sibling wildfire-service-territory-overlap decided this shape in advance for its own second retrieval (ADR 0012: paired, never modelled) and the same rule fits here: two pins invite a story about improvement, and the page must refuse to tell it. The build stays deterministic; history artifacts are inputs committed under site/, and the page is rendered from them like the other two.

Scope

• History directory and a loader that refuses an artifact whose source.retrieved or sha256 disagrees with its filename or, for the current pin, with sources.py.
• render.py: history.html with the same palette, markers, WCAG gates, and prose-number tracing test as the other pages.
• An ADR recording the paired-difference rule and the refusal of trend language, with a page test that greps for it.
• tests/test_published_site.py extended to the new page; site-offline builds it from two fixture pins.

Out of scope

• Any per-year or per-incident rate of change.
• Retrieval scheduling.

Done when

• With two fixture pins, the page lists each changed per-year count with both values and the difference, and an incident present only in the later pin as "not in earlier retrieval".
• With one pin, the page says there is one retrieval and no comparison, rather than an empty table.
• A history artifact whose hash disagrees with its filename fails the build.
• The page passes every existing page gate.

Pointers

• src/perimeter/render.py, artifacts.py, sources.py (RETRIEVED), site/data/, docs/adr/0009-the-clock-lives-in-the-gate.md
• wildfire-service-territory-overlap/docs/adr/0012-two-boundary-retrievals-are-compared-paired-never-modelled.md

Proposed with AI assistance.

[ChelseaKR]

Triage against origin/main @ d4f533f (2026-09-06). Not built: no site/data/history/, no history.html, no loader.

Buildable, with one honest caveat about what it can be tested against. There is exactly one pinned retrieval (2026-08-07), so every "Done when" row except one exercises fixture pins rather than real ones. That is fine — and the issue anticipates it: "With one pin, the page says there is one retrieval and no comparison, rather than an empty table." That single row is the most important test in the issue, because an empty comparison table published as if a comparison had been made is precisely the defect class this repository organises against.

The design constraint to keep, and the reason it will be tempting to break: paired differences, both values shown, no line fitted, no direction stated. Two pins invite a story about improvement and the page has to refuse to tell it. An ADR plus a page test that greps for the refusal, as the issue asks, is the right enforcement — prose alone drifts.

One thing that needs settling before it ships and is not a coding decision: history artifacts become committed inputs under site/, which means the repository starts retaining every past artifact permanently. That is small in bytes and is still a retention choice. Worth an explicit yes rather than an inferred one.

No owner decision blocks writing the loader, the refusal, and the fixture-pin tests.

Triaged with AI assistance.

[ChelseaKR]

Triage, 2026-09-07. Unbuilt, and blocked on something no code supplies: a second pin.

There is one retrieval (2026-08-07). A ledger of paired differences between consecutive pins has, today, exactly one pin and no pair. The one-pin case ("with one pin it says there is one retrieval and no comparison") is buildable against fixtures; the case the page exists for is not, and shipping only the empty case is apparatus with nothing to compare.

Worth distinguishing from what already landed: #71 shipped src/perimeter/diff.py and make diff, which compares two artifacts leaf by leaf on demand. That is the once. This issue is the every-time: committed history under site/data/history/ and a rendered site/history.html. Neither exists on main, so this is genuinely open.

The sibling wildfire-service-territory-overlap #97 is the same shape and is blocked the same way.

Prepared with AI assistance; reviewed before submission.