fix: regressions in the sweep surfaced by adversarial code review

[Xiddoc]

Summary

An adversarial code-review workflow over the two merged sweep PRs (#248, #249) found four high/med regressions that passed CI at 100% coverage because the existing tests had blind spots. This PR fixes them and closes the test gaps.

HIGH — the default binder: vm boot was broken (#162 / #174 / #176)

QemuProcess._pid_is_qemu() matched str(instance_dir) in /proc/<pid>/cmdline, but the default plain up argv (boot_cache: false, the shipped default) resolves kernel/rootfs from a shared artifacts cache outside the instance dir and references nothing under it. So is_running() was always False on the plain path, which:

• made _wait_for_adb_connect() abort instantly on a perfectly healthy boot (Warm boot_cache resume has no cold-boot fallback and never polls is_running() during the adb wait, so an unrestorable snapshot burns the full timeout then misreports TCG slowness #176), and
• made terminate() orphan the QEMU instead of killing it (VM up()/_up_cached() leak the running QEMU process when the adb-connect wait times out, orphaning the emulator and trapping the next up #174).

The e2e tier missed it because it forces boot_cache: true (where the overlay/monitor do live under the instance dir). Fix: key identity on the instance-unique hostfwd ADB port, which every launch argv carries; the config-gone orphan teardown falls back to a best-effort qemu-system check. Tests now drive a real build_qemu_argv with paths outside the instance dir.

HIGH — frida accepted a truncated download (#228)

The incremental LZMA loop never checked decompressor.eof, so a stream that ended before its end-of-stream marker (a truncated download) was silently accepted as a complete binary — the one-shot lzma.decompress it replaced would have raised. Now rejects truncation, and caps per-call output (max_length) so a single compressed chunk can't expand without bound.

MED — gapps_vendor override note re-emitted per construction (#220)

The note was an Android model_validator firing on every construction, so a fleet scan reprinted it once per pinned-vendor instance — the repeated-note anti-pattern #202 had just fixed for its sibling. Moved to load_yaml, deduped per resolved path.

MED — ls/modes off-TTY table still folded wide cells (#204)

rich clamps a print(width=…) request to the console width (80 off-TTY), so a load-bearing cell wider than 80 was folded across lines. The render width is now set on the console so it stays on one physical line.

Also

• Completes module.sha256 shape validation (frida.sha256 / module.sha256 accept any string (no 64-char-hex check), so a fat-fingered digest fails late with a misleading hostile-mirror message #194 — the half deferred in fix: config validation, download hardening, builder robustness & isolated bug sweep #249) and fixes the sha256 regex to reject a trailing newline (fullmatch, not match).

Testing

Full local gate green: ruff, ruff format --check, mypy --strict (src+tests), pytest -p no:randomly at 100% line+branch coverage (1837 passed, 6 docker-daemon skips), shellcheck/shfmt, codespell, changelog lint. New tests drive the real plain-path argv, a truncated .xz, the per-path note dedup, the >80-col table cell, and module.sha256 rejection.

🤖 Generated with Claude Code

https://claude.ai/code/session_01P1YsNMpGizhSDDPCBu5cut

---

Generated by Claude Code