An end-to-end drive of the update flow against a local HTTP server hit
two Windows-only tar quirks that a first-cut MVP wouldn't catch:
1. Git-Bash tar (MSYS2), which comes first on PATH when Git-for-Windows
is installed, treats drive-letter paths as `host:file` remote-archive
syntax. Sidestepped with --force-local (also silently accepted by
Win10's built-in bsdtar and by GNU tar).
2. Even with --force-local, MSYS2's argv-conversion layer mangles
backslashes in Windows paths, so `C:\Users\...\tmp\dir` arrives at
tar as `C:\Users...\dir` and it can't open the path. Passing
forward-slash paths (`C:/Users/.../tmp/dir`) dodges the mangler;
bsdtar and GNU tar accept them as-is.
3. sign-addon-update.mjs was tar'ing the addon directory as a subfolder
(`screenshot/addon.json` inside the archive), so the client
extracted to `<tmp>/screenshot/` and then failed the id+version
re-check because addon.json wasn't at the root. Now the signer
tars the CONTENTS of the addon dir via `tar -C <addon-dir> .`, so
entries live at the archive root where the client expects them.
All three surfaced from `scratchpad/decoupling-test/run-test.mjs`, which
now walks the full path — sign, serve, fetch, verify, download,
extract, stage, promote, backup — plus three signature-tamper negatives
and the empty-pubkey short-circuit. 15/15 checks pass.
95d199c2f2