From 7b46dbcc3ecf08ad33b43c998072bfee4e6350e8 Mon Sep 17 00:00:00 2001 From: kisfenyo Date: Sun, 15 Nov 2026 16:50:56 +0100 Subject: [PATCH] for the release of release 4: two ways to prove the off-site deletion on a test store, with a pick Claude-Session: https://claude.ai/code/session_0159rPz1ZhFKsS53msqPYxtS --- .../E/proving-the-deletion-on-a-test-store.md | 78 +++++++++++++++++++ 1 file changed, 78 insertions(+) create mode 100644 documentation/audits/build-release-4b-2026-11-15/E/proving-the-deletion-on-a-test-store.md diff --git a/documentation/audits/build-release-4b-2026-11-15/E/proving-the-deletion-on-a-test-store.md b/documentation/audits/build-release-4b-2026-11-15/E/proving-the-deletion-on-a-test-store.md new file mode 100644 index 00000000..b5e9773d --- /dev/null +++ b/documentation/audits/build-release-4b-2026-11-15/E/proving-the-deletion-on-a-test-store.md @@ -0,0 +1,78 @@ +# How to prove the off-site deletion of a removed app on a test store — two ways (2026-11-15) + +For the release session of release 4, Part A. **Never on a household's store.** The test household is the scratch +guest's own (guest 9202 on demo-hp, with its own off-site store); the app is a throwaway app that was installed, backed +up off-site at least twice, and removed. + +What must be seen in the end, in both ways: the mail to the test household's address with the day; the page's marked +line; one take-back that works; after the second press, one weekly window whose answer names the ended app's +snapshots (`removed_for_ended_apps`); the files gone at the store; a LIVING app's snapshots all still there (counted +before and after); the next 07:10 check silent; the page's history line. + +Three clocks stand between a press and a deletion — all the hub's own, none the box's: + +| Clock | Where | What it holds | +|---|---|---| +| 30 days of absence | hub table `offsite_app_presence` | when the hub first saw the app missing from the box's app list | +| 30 days after the notice | hub table `offsite_end_notices` | when the mail service accepted the notice | +| the box's own record | the scratch guest's `settings.json`, `offsite_ends` | removed at, ends at (the box will not offer an app before its own `ends_at`) | + +## Way (a) — by hand: the three dates are set back 30 days + +| Step | What is changed by hand | The copy made first | Who must say yes | +|---|---|---|---| +| 1. Install the app, two off-site backups on two days (or two „run now”), remove it with „delete backup data”, read the mail and the page, take it back, press again | nothing by hand | — | — | +| 2. Wait for the next report (the hub then holds the two rows) | nothing | — | — | +| 3. On the live hub: the two rows of the TEST household only get dates 31 days back (`offsite_app_presence.absent_since`, `offsite_end_notices.told_at`) | **two rows of the live hub's database, written by hand** — the first hand write to that database since the project began | the hub's own snapshot job, run once before it (a whole copy on the hub's volume, and its push to ep0) — and the two rows saved as text | **the operator**: it is a write to the live hub's database, on DooPlex; CC does not do it alone | +| 4. On the scratch guest: the record's two dates set 31 days back in `settings.json`, with the controller stopped for the write | one file on the scratch guest | a copy of the file beside it | CC (the scratch guest is a test box) | +| 5. Open one clean-up window for the test household by the operator's one-shot grant, on the day | the window is the product's own | — | the operator (the grant is his button) | +| 6. Read: the window's answer, the store's listing before and after, the 07:10 check next morning | reads | the listing before, saved | — | + +*What it proves:* the whole path after the 30 days — the agreement, the allowance beside the cap, the removal by id, +the audit's explanation, the pages. +*What it does not prove:* that the three clocks count 30 days by themselves (they were set by hand); the notice's +retry; anything about a household whose mail does not go. And the hub's rows are no longer what the hub wrote — the +test household's rows must be deleted after the proof, or a later reader takes them for real. +*Cost and risk:* a hand write into the live hub's database. A wrong `WHERE` there touches a real household. + +## Way (b) — by time: the scratch guest presses on the release day + +| Step | What is changed by hand | Who must say yes | +|---|---|---| +| 1. On the release day: install the throwaway app on the scratch guest, two off-site backups, remove it with „delete backup data”; read the mail, the page; one take-back; press again | nothing | — | +| 2. Day 1 to day 30: at every weekly window read the window's answer for the test household — it must REFUSE the app and name the reason („told on … — N full days, 30 needed”) | nothing | — | +| 3. The first window on or after day 30: read the answer, the listing before and after, the next 07:10 check | nothing | — | + +*What it proves:* everything way (a) proves, AND the three clocks, AND 30 days of refusals with the right reason. +*What it does not prove before day 30:* the removal itself. Until then the feature is released with its last step +never seen on a real store. +*The cost:* the test box is the first box in the world whose copies this feature deletes — which is the point — but a +household that presses in the same weeks is only a few days behind it. + +**What protects a household that presses in the same weeks (way b):** + +1. Nothing of a household's is removed before ITS 30 days after ITS mail. A household that presses on the release + day is due on the same day as the test box — so the test box must press FIRST: the release session presses on the + scratch guest before the controller release reaches any other box (the per-box floor of the scratch guest and the + test boxes first, the press, then the global floor a day later). The test box is then at least one weekly window + ahead of every household. +2. The mail tells the household the day and how to take it back; the operator gets a line for an app the hub never + saw, and for a removal of more than half of a store. +3. If the test box's first window goes wrong, the operator has one switch per household: the weekly windows off + (no window, no removal). **To write before day 30:** the exact place of that switch, and that the release session + tried it on the test household. +4. One window removes at most the cap's size of a removed app's copies; the rest waits a week. + +## CC's pick: (b), and (a) only if the operator wants the removal seen before any household can be due + +Why: (b) changes nothing by hand and proves the clocks, which are the whole protection. (a) writes into the live +hub's database by hand, to prove the part that the tests already drive against a store that really forgets +snapshots. The honest cost of (b): for 30 days after the release the last step is proven by tests only — and the +release note must say so. + +A third thing that needs no box and should be done in any case, in the release session: the 30 days of both hub +clocks and the box's clock **as tests with a moved clock** already exist (`TestD72_*`, `TestD84_*`, the notice-lock +tests); read their list against the eight „must hold” lines of the brief once more before the deploy. + +**Not done tonight:** no hub row was read or changed; no box was read; the names of the two hub columns above are +from the source of 2026-11-15 and must be read again at the proof.