Fix #27's zip link, stub probe and divider margin; pin engine .27

[scottjones]

Fixes the three defects from @maralcbr's post-merge review of #27, in his priority order. I was wrong earlier in this description, where I said the stub and divider fixes were already on main.

1. Download link → zip (977e588). Since #38 the installer is published only as installer/<version>/<INSTALLER_FILE_STEM>-<version>.zip, but both release templates still gave …/Omarchy-MX-Mac-Installer-2.1.0.pkg as the update link. The identity test pinned the same .pkg name, so it didn't catch this. Both templates and the test now use .zip.

2. Stub containers no longer probed (876df11). A real stub reports the version of the macOS it was made from, so any(os.version …) still sent stubs to the diskutil limits probe. On a Mac with an existing Asahi or Omarchy install, one failed probe failed the whole inventory. The filter is now any(os.version and not os.stub …). The test's fake stub now carries a version and stub=True, as upstream's OSInfo does, and the test fails against the old filter.

Engine rebuilt and re-pinned as .27: installer-v0.9.2-omarchy.27.tar.gz, 17,839,422 bytes, SHA-256 4f9241b0139ba6ccdcfdb3484831002e07e36ba50274279c641ca2020593a15a. Two repacks with macOS /usr/bin/python3 3.9.6 produced identical bytes. Compared with .20, only omarchy_runtime.py and version.tag changed. verify-archive-modes and verify-source-lock pass. The app's bundled inspection pin, the packager and both templates select .27, and docs/extraction.md records it.

3. Divider can't freeze (4820d04). InstallerAllocationRecommendation adopted the recommended (doubled) size as the minimum before withholding the drift margin. In a band of free-space states this made min equal to max, with no margin left. The margin is now subtracted first, and the doubled size is adopted only strictly below the margin ceiling. New tests sweep that band in quarter-GiB steps, checking that the divider has a range and the margin survives; both fail on the old code. Near the partition floor the existing best-effort clamp is unchanged.

Validation (macOS, M4, Xcode toolchain): ./test/all passes (Homebrew Bash 5, Python 3.13). swift-format lint --strict is clean. swift test passes in debug and release: 519 tests, 1 skipped, 0 failures.

Not in this PR: publishing .27 and the signed catalog, and the hardware test (smallest-size encrypted install, reboot, full update). The "Later" items are also not included.

Effect on #39: once this merges, I'll rebase #39 and rebuild its engine once, as .28, on top of .27.

🤖 Generated with Claude Code

[maralcbr]

Reviewed at 9787a83. All three #27 follow-ups are fixed: the templates and identity test use the published zip path, stubs are no longer probed (with a realistic fixture that fails on the old filter), and the doubled size is adopted only below the margin ceiling, so the divider keeps a range and the drift margin survives. I rebuilt .27 twice with /usr/bin/python3 3.9.6 from the locked v0.9.2 checkout and got exactly 4f9241b0…, 17,839,422 bytes; against a .20 rebuild from main only omarchy_runtime.py and version.tag differ, and verify-archive-modes and verify-source-lock pass. ./test/all and swift test pass locally (677, 1 skipped).