Apply E2E review fixes that missed the #22 merge window

[oratis]

PR #22 was merged at 05:50:48Z, four seconds before the review-fix commit ed74577 landed on its branch — so the fixes from the deep-review comment on #22 never reached main. This PR carries exactly that one commit.

Contents (see the review comment for full rationale):

1. Reachable preflight failure gate — the ANDROMEDA_INSTALLER_PREFLIGHT_FAILED branch was unreachable: %pre --erroronfail aborts before partitioning, so the partition-probe gate exited first with a misleading harness-timeout message. The failure-detection block now runs before the partition probe, and ESP diagnostics presence is a third independent trigger.
2. Honest matrix budget — the "verifier short-circuits at state=complete" comment was false (andromeda-ci-verify:75 runs daily-driver-verify unconditionally inside the 12-min unit timeout). Comments now state the 600 s host cap is deliberately below the unit timeout, with the rationale; budget arithmetic refreshed (prep ~6 min per run 30732275955; total 176 < 180).
3. Occurrence counting — grep -c counts lines, not occurrences; two markers collapsed onto one line after CR-stripping would satisfy "exactly once". Now grep -o | wc -l.
4. Marker/assert lib cleanup — dead serial_grep deleted; input-redirection stderr leak guarded; require_marker_fixed (fixed-strings) added and used for interpolated scenario markers; andromeda-ci-verify.service budget comment names its ordered predecessors; matrix artifact glob narrowed to logs/text (no more qcow2/OVMF uploads).

Validation: bash -n clean on all changed scripts; test-installer-platform-guard.sh pass (14 cases); layer-budget count mode pass; functional smoke tests for the counting and normalization changes.

🤖 Generated with Claude Code