MOB-183: mob.deploy verifies release bootfile on Android, not just ERTS

[GenericJam]

Summary

• Non-native mix mob.deploy on Android used to check for otp/erts-*/bin/erl_child_setup and pass when it was present — but the OTP release directory could still be missing, in which case the emulator loaded, the app crash-dumped at boot with {'cannot get bootfile', .../otp/releases/29/start_clean.boot}, and the deploy had reported success. Third sighting of this failure family (mob_dev#54).
• Probe now verifies BOTH erts-*/bin/erl_child_setup AND releases/*/start_clean.boot in one run-as-wrapped ls. Classifier extracted to public MobDev.Deployer.classify_android_runtime_ls/3 — matrix testable without a device (same pattern as parse_devices_output/1, parse_simctl_json/1).
• Recognises three ls: <glob>: No such file variants (Toybox default, older Toybox not found, GNU coreutils cannot access). Anchors on the exact glob string so ERTS-missing vs bootfile-missing don't confuse each other.
• Fixed the pre-review revision's silent-pass: modern adb-shell (v2, default since Android 7) forwards the inner shell's non-zero exit as the session code, so a partial-runtime device came back as {:error, out} from run_adb and skipped the classifier entirely. The classifier now runs on out regardless of exit status; empty-out handles the true adb transport-failure case as best-effort :ok.

Test plan

• MobDev.Deployer.classify_android_runtime_ls/3 — 8 tests covering: both present → :ok; run-as: prefix → run_as_unavailable; both missing → erts_missing (preferred over bootfile in the both-missing case); ERTS present + bootfile missing → bootfile_missing; Toybox not found; GNU cannot access; the MOB-183 regression fixture (adb-shell error arm with real ls output); empty/nil output → :ok (transport-failure fallback).
• mix test full suite — 2431 passed / 10 excluded (integration).
• mix credo --strict — clean.
• Adversarial pre-commit review by fresh subagent found the classifier-never-runs blocking issue; fix applied and specifically regressed with a fixture-based test.
• iOS follow-up: the sim variant of this failure family lives under ~/.mob/runtime/ios-sim/<app>/ (host-side, not device-side). Separate ticket.

Linear: MOB-183
GH: closes mob_dev#54

🤖 Generated with Claude Code

https://claude.ai/code/session_01Ezeu2FEZfRajeoL4ywEUDJ

[GenericJam]

Rebased onto current origin/master (52b0a72 — MOB-169). This branch had been cut from 21ba7f0 (0.7.1 bump); MOB-169 landed in between and my "check for parallel agent work" only looked at open PRs, not just-merged commits since the base. Rebase was clean — MOB-169 touches deployer.ex but nowhere near ensure_erts_on_device, so no code-level interaction.

Post-rebase gates:

• mix compile --warnings-as-errors — clean
• mix test — 2455 passed, 10 excluded (integration)
• mix credo --strict — clean

Sibling PR #78 (MOB-182 — refuse ambiguous multi-device deploy) was closed as obsoleted by MOB-169's stronger --all-devices / --all-physical gate. This PR (MOB-183 — runtime bootfile check) is unrelated to that surface and stands on its own.