[janicduplessis]

Description

npm 1.20.2's version commit (700cb25, #3428) was not tagged: the full Windows unit suite failed on it (#3437, fixed by #3443). Main has since gained several package changes that will ship in the tag, so docs/releases/1.20.2.md must list them first.

Solution

Adds the package changes merged after 700cb25: the cache-registration guard (#3434) and the remote server-update downgrade refusal with corrected build-grant docs (#3432), both under Security; stim-server build upload limits and cleanup (#3350, with #3439's abort on unreadable mirror records); stim macos waiting for LaunchServices (#3435), which replaces the manual lsappinfo docs line from #3413; Android relaunch with am start -S (#3327); and the 355 s hosted install wait (#3355). #3441 and #3443 are type and test fixes with no user-visible effect, and #3433, #3386 and #3427 are outside the npm packages. The QA counts are refreshed.

Test plan

node scripts/release-qa-matrix.mjs v1.20.1 reports 250 changed paths, 87 mapped, 0 unclassified, matching the notes; format:check passes. After merge, dispatched CI (including windows-full) runs on the merge commit before the tag.

[janicduplessis]

Fresh review of eb6a64e. Not approving yet: one entry misdescribes the old behavior.

Verified against the diffs (SHAs, PR numbers, behavior, bounds all match): a6c9ba3, fa16761, fdae568 plus its follow-ups f87b611 and 529b901 (the latter two are type-only/mirror-record validation; the 'pruning stops when a mirror record cannot be read' claim holds), 0d5bbf7 (20 s wait, warn-and-return, --remote skip, guide/website text), b5c6b9d (INSTALL_TIMEOUT_MS 300 s to 355 s; prepare, session and offer waits compute to their old values). a3a4c9f predates 700cb25, so dropping its Docs line in favor of the 0d5bbf7 line is right. git log 700cb250f..origin/main -- packages lists nine commits; eight are covered, the ninth (84fb434) is a test-only change. QA counts reproduce: release-qa-matrix.mjs v1.20.1 gives 250 changed paths, 87 mapped, 0 unclassified, 8 of 8 rows.

Finding (313f2c1 / #3327): the entry says the new am start -S launch means 'a slow adb no longer refuses the relaunch'. Before #3327 the unknown-state case did not refuse. stopRunningAndroidApp treated an unreadable process state as not running (if (!pid) return {}), so Stim launched without stopping and an already running app kept its old process and bundle. The refusal-on-failed-stop behavior is the known-pid path, which is unchanged (and the iOS #3308 background). Suggested wording: 'When Stim cannot read an Android app's process state, a relaunch now launches with am start -S, which force-stops the app as part of the start; before, it launched without stopping, so a running app could keep its old process and bundle. An app without a launcher activity gets a best-effort am force-stop before monkey. A known running pid still gets a separate force-stop, and a failed one still refuses the launch.' Also note #3327's own test plan says the real-tool am start -S check was not run; not claimed in the notes, just flagging it.

Minor, optional: b5c6b9d also makes the hosted client's waits cancellable on run cancellation; the entry does not mention it, which is fine for a refactor line.

needs-human: not applicable. The diff is a notes-only change, and the listed source changes are hardening or bug fixes that do not widen exposure, deletion or secret access.

I did not add review: ai-approved or mark ready. Re-review after the wording fix.

[janicduplessis]

Re-review at 96c8273: the only change since eb6a64e is the 313f2c1 line, and it now matches the code (unknown state launched without stopping before; known-pid force-stop still refuses on failure). All other entries, SHAs and QA counts were verified in the earlier comment and are unchanged. No actionable findings; review: needs-human does not apply.