Heretic II: payload-free observation and IAT reentry

2026-11-12. Evidence: scratch/runs/20261112-heretic2-payload-free-iat-observer-2245/.

The parent WriteFile decoder violated the restriction. Its decoded output was not inspected or reused here; prior compliance is not claimed. The archived parent box and its historical observations remain untouched.

Before this original, both composed API hooks were replaced with metadata-only recorders. They read API-name metadata, public stack arguments and return address, the transferred-count DWORD and existing runtime getters. They do not dereference file buffers, decode guest strings passed as arguments, inspect guest private structures, call guest functions or query GetLastError/GL. The renderer observer's additional GL queries were also removed. No runtime or guest patch was made.

The source-backed synthetic fixture executes a real x86 CALL through an IAT into the frozen 8a834 dispatcher and lazy VFS. The old classifier produces UNKNOWN because entry EIP is the caller while io_block sets post-handler EIP to the reentry thunk. The new classifier reads the existing com_callback_get_thunk getter and requires matching thunk, status997, yield12/flag1, API-call count, thread, stack frame, public return address/arguments and filesystem import arguments/thread. Completion retains guestReturn UNKNOWN until an independently observed runtime-return boundary matches. Historical parks were not relabeled.

Actual composed legacy/new hooks plus full frozen owning-Worker startup pass Node tests with payload/private-address reads forbidden; the native Chromium Worker independently reproduces the IAT failure and fix. Synthetic COM entry, status995 cancellation, identical throws, once-only forwarding, receiver/argument and return identity, missing-state UNKNOWN and getter/sink isolation pass.

Both recorders use bounded 4,096-row selected rings and bounded pending/frame state. Drawing and unnamed noise is counted and suppressed. Persistence uses bounded snapshots rather than per-call append logs. Overflow and phase tests pass before allocation. The driver resets and phase-tags both buffers immediately before the authorized load Enter, preserving startup counts separately. Retained read requests have a 30-second hard cap, with UNKNOWN and no replay on expiry.

Exact four-save closure, actual provider validation, seven negative stage cases, and final staged imports pass. The locked build passes and produces unchanged module 04c534192ee3fe39e388536c2ce710c12c47a5cd2ca4d95c19f9d00b1e8f0a17. The baseline is 8a834a3dc7affc8c6e53245beab56e87a32b81e8; renderer provider is 4235d20d7fc2404b38f6e4f7f80c92d5f94f1513b4d6d7dcf04be34162a4b087. Unchanged identity/freshness/independent-UP proofs are reused by their SHA.

A fresh observable independent two-hour watchdog preceded creation of bx_6c3vkfte at 22:40:51Z. All 2,460 remote hashes and actual provisioning join 0/both EOFs passed. The first local original-launch check refused below the 1,500,000,000-byte floor without submitting a launch. A bounded local hold on the designated recovery handle then observed actual free space above the floor; no cache or other worker's files were changed. One original was then launched.

The actual original reached normal Game → Easy → upright textured docks HUD100, then mounted and hash-read back exactly all four saves. The console blank prompt was reviewed. Exactly load crate0400 was typed once. Its first capture still showed a blank prompt; a necessary read-only settled capture then showed exactly the command. Enter was never sent; no load or Exit8 occurred in this run.

This exposed a new driver blocker I introduced: the phase-reset gate asserts that the latest capture label equals load-typed. A newer valid typed-settled capture therefore prevents reset/Enter. test-driver-phase.cjs executes the exact frozen driver function with mocked host boundaries: load-typed permits one reset, DOWN and UP; typed-settled rejects before reset or input. Its recorded source SHA matches the frozen original helper. I should have included this actual-driver path before freezing. The earlier phase/overflow tests covered the composed recorders, not this label coupling. No live helper was patched, no Enter was submitted to a known rejecting gate, and no second original was started.

The recovered original startup rings independently qualify 149 actual caller-to-thunk parks and 149 correlated retries in each recorder. Each holds 4,096 selected rows, with 14,957 evictions, zero observer/sink errors and no drawing rows. The retained ReadFile posts comprise 149 internal parks and 414 completed handler frames; all 563 guestReturn fields remain UNKNOWN. Counts explicitly remain startup phase with zero resets. This is actual startup correlation evidence, not validation of retention through an ordinary load. That load remains untested. No historical observation was backfilled. No file payload was decoded by these hooks; the mount readback was solely the authorized host-VFS hash comparison.

Ordinary browser closure was ACKed at 22:57:35Z, not a guest quit or ExitProcess. Chrome exited 0; the actual original owner wait joined 0 with stdout/stderr EOFs, and its confirmed-dead profile was removed. The retained observer read completed without the parent's disposal hang. 134 unique blobs / 164 paths / 6,607,591 gzip bytes were recovered with remote raw SHA and complete local roundtrip checks before delegated normal stop. bx_6c3vkfte was archived at 22:58:53Z, expiry null, receipt stop_3af57d842f39ffab, completed/unforced/error null. Normal stop and independent watchdog actual waits are 0 with both EOFs. The local exec session for start.py ended 143; no manager-parent wait record exists, so that wait/EOF result is UNKNOWN. No billing box remains for this task.

Only task-owned completed logs/proofs were losslessly compressed during the local holds. The designated recovery handle and its files were untouched. No RAM, resource arithmetic, reserve/cache change, guest RE/disassembly/private-structure inspection, extraction, guest patch, substitute DLL, added GL/GetLastError query, new clock/config write, child worker, other game or public deployment was introduced. The inherited threads=1 launcher setting remains part of the unchanged route. There is no new SW31 pair, Exit8-cause, full27, FPS, audio, parity, playability or general crash claim.

Git plus binary diffs and owned helper patches are the recovery archive; media remain path/size/SHA references. RECOVERY.md explains compressed proof segments. The next original requires a meaningful driver regression that allows a newer reviewed bound settled frame after the sole typing intent, retains all freshness and identity guards, and resets both recorders exactly once before Enter. The existing frozen helper must not be reused for that path. This task stops at the new distinct pre-Enter driver blocker; Exit8 cause remains UNKNOWN.