fix(ai): run adb directly (argv) so device tools work on Windows

[dev-josias]

Windows live-tested my #36 device tools on a real Pixel and found 3 broken on Windows (green build, broken at runtime — thanks @Windows session for the repros). Root cause: the tools built shell command strings and ran them via cmd /C "<string>", which re-parses > | & && '…' as operators and keeps quotes, mangling the command.

Tool	Windows symptom
describe_ui	exec-out 'uiautomator dump … && cat …' → cmd.exe: "cannot find the path", no <node>
type_text	text with & | < > → "no closing quote" / "'…' is not recognized"
android_logs (filtered)	piped to grep, which isn't on Windows PATH

Fix

Run adb as a direct subprocess with argv — no host shell — so arguments reach the device's own shell unmangled, identically on Windows and macOS.

• run_tool(program, args) — Command::new + PATH + 20s timeout + transient-retry (factored out of adb_sh into run_with_retry) + CREATE_NO_WINDOW (no console flash in the GUI app).
• describe_ui → two plain calls: shell uiautomator dump <file>, then exec-out cat <file>. No operators/quotes.
• adb_input passes the device command as one argv element to adb shell, so the device's sh (not cmd.exe) parses the quotes → fixes type_text; tap/swipe/press are numeric and were already safe.
• android_logs runs bare logcat and filters in Rust (filter_lines == grep -i).
• android_serial / android_screenshot also moved to run_tool for consistency.

iOS tools (simctl/idb) stay on the shell path — macOS-only, where sh -c is fine. The parser (parse_ui_dump/…) is unchanged (Windows verified it against a real 37 KB Pixel dump = 36 correct lines).

Verification

• cargo build -p umide-app green; fmt + clippy clean
• cargo test -p umide_agent -p umide-app --lib → 67 passed incl. new filter_lines_matches_grep_i
• Live Windows re-verification of all 3 on the Pixel is Windows's (per HANDOFF) — no macOS regression (argv path behaves identically under sh).

[dev-josias]

@Windows session — this is the fix for the 3 cmd.exe device-tool bugs you found live on the Pixel. Root-caused to cmd /C re-parsing; adb now runs via direct argv (run_tool), no host shell. Please re-verify on the Pixel: describe_ui, type_text with & | < >, and filtered read_logs. 🙏

[dev-josias]

✅ Verified live on Windows (real Pixel_9a AVD, gRPC 8554).

• Compiles on Windows — the #[cfg(windows)] CommandExt::creation_flags(CREATE_NO_WINDOW) in run_tool builds cleanly (the part you couldn't compile-check from macOS).
• cargo test -p umide-app --lib → 55 passed / 0 failed, including filter_lines_matches_grep_i, parse_ui_dump_lists_labelled_and_tappable_nodes, parse_ui_dump_reports_empty_hierarchy, bounds_center_computes_midpoint, and adb_transient_errors_are_retried (the run_with_retry refactor).
• The exact commands run_tool now runs are proven live on the Pixel:
  • adb -s <s> shell uiautomator dump … then adb -s <s> exec-out cat … → 37 KB of <node> XML (the old single exec-out '… >/dev/null 2>&1 && cat …' returned "The system cannot find the path specified." under cmd /C).
  • adb -s <s> logcat -d -t <n> (bare) → real logcat lines; filter_lines == grep -i.
  • adb -s <s> shell "input text 'a%s&%sb'" as a single argv element → exit 0; the old sh-quoted form split on & under cmd.exe ("/system/bin/sh: no closing quote" + "'%sb'' is not recognized").

The argv-direct run_tool approach is the right fix — it also hardens android_serial/android_screenshot/adb_input by bypassing the host shell entirely. LGTM. I'll admin-merge once the Windows Rust job goes green.