scripts: read the install/deletion cohort, the report that made v1.49 not ship pings

[JasonYeYuhe]

Why

App Store Installation and Deletion Standard carries an App Download Date on every row, making it the only Apple feed that yields a real cohort: how long after downloading did this install get removed.

That report is the stated reason v1.49 shipped a first-run window instead of D1/D7/D30 retention pings — the pings would have been conditioned on the user having already opened the popover and accepted the disclosure card, i.e. on having cleared the very hurdle under investigation, so they could only ever have described the survivors.

But acquisition_reconcile.py didn't read it. The justification was one session away from evaporating. Now it does.

Also adds App Store Purchases Standard, short enough to print in full: one purchase, 2026-06-24, iPhone / CN / Alipay, $1.03 proceeds.

The coverage line caught its own bug, which is the point

The first version compared total Install events against first-time downloads and printed iPad coverage of 114 % — impossible for a subset, and visible on the first run. Install spans every Download Type (first-time, redownload, auto-download, manual update); first-time downloads are one of them.

Fixed to compare like with like, which moves the sample fraction from a comfortable-looking 18–64 % to the real figure:

device	installs	(first-time)	deletes	first-time coverage
Desktop	19	5	6	5/108 = 5 %
iPhone	129	25	20	25/203 = 12 %
iPad	8	1	2	1/7 = 14 %

That is not a cosmetic fix. This feed is a consenting-user sample, not a census, and a far smaller one than the first version implied. The output now says so in the place someone would actually read it:

the SHAPE of the latency distribution is usable;
deletes / installs is NOT a deletion rate and must not be quoted as one — it divides one population by another.

Current shape

n=27 — median 0 days, 63 % same-day, 74 % within one day, 89 % within seven, 100 % within thirty.

A median at or near zero is the signature of "opened it, didn't get it" — abandonment before the app ever proved itself, which no in-app counter can observe because the counter only fires after the menu is opened and the disclosure accepted.

Track this series across releases: it is the readout for the v1.49 first-run window, and the honest one, because it is measured outside the app rather than by the app.

Verification

• --self-test still passes (the anti-threshold guard is applied to the new feed too).
• Run end-to-end against live ASC; output above is real.
• ruff clean — and as of ci: make ruff actually lint the Python it claimed to lint #437 that now actually runs on scripts/.
• The plan's own caveat has been corrected in cli-pulse-internal (41d4965), since it quoted the same inflated 18 % figure.

🤖 Generated with Claude Code

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.