RichEdit native callback fixture technical review — 2026-10-23
Historical source last updated, 2026-10-23. Runtime and implementation claims below describe that dated source; see current app notes and source for subsequent changes.
Verdict: accept the corrected fixture for pending native qualification. Commit 233a627d6a1ece2a95fb5600d000aa96be5c9e17 had a precise assertion
gap; this review supplies a narrow test correction. Native execution remains
UNEXECUTED. This is neither full-module validation nor runtime acceptance.
Inputs: published task/rich-edit-native-callback-fixture-oct23 at 233a627,
remote main bb9c496ed13095cc832fdf607f0937fb9c14b7de, and accepted source
3c01665cdeb65ca7aebfdae7aec8dba4c3499884. Strict authenticated ls-remote
verified both published inputs. The edit fragment is the identical blob
6e07d34aedf6199a419fcbe0bd79b7b6a409dd8f in accepted source and main. All 63 donor artifact SHA256 seals verified; the donor remains unchanged.
Defect and narrow correction
The original 70-sibling oracle accepted only aggregate LRESULT 210 and callback
count 140. Replacing two nominal calls by results [0,6] and counts [1,3]
preserves both totals. The independent replay demonstrates that the original
oracle accepts these observations. Every sibling now checks result 3, exactly
two callbacks, error 0, visible text OWN, stack canary and interrupted state
before contributing to totals. Compensating failures reject individually.
The original PE checker also accepted a sliced loop shortened to one iteration and deleted timeout spin instructions. It now requires the emitted 1,000,100 iteration decrement/backedge sequence and unconditional self-loop tails in both timeout callback locations. Native cases invoke this PE check before loading the fixture. This prevents those specific scenario substitutions from qualifying through otherwise identical result/error observations. These are structural guards, not measured execution-duration evidence.
Reviewed paths and oracles
- Ordinary: public PE CreateWindowExA imports create STATIC and two
RichEdit20A controls. Outer
send_messageand nested SendMessageA IAT calls reach the existing WAT native control path; no private wrapper or fabricated owner record is injected. Data/EOF counts, result, dwError, visible text and callback argument observations prevent the normal no-callback/empty success. - Nested and nested-timeout: callback snapshots preserve six non-EAX/non-ESP registers plus ESP; EAX is checked as the public nested result. Inner text, count/error, canary and outer completion must agree. Full supported outer register/control state is collected before sibling tests overwrite counters.
- Sliced/timeout: emitted long decrement loop and infinite self-loops are independently decoded with objdump. Accepted source repeatedly runs one million blocks, checks retained child/root/depth before retirement even on timeout, and caps at 64 rounds. Offline inspection cannot prove these budgets execute correctly or measure the actual number of runtime rounds.
- Capacity: callback depth 63, trap
unreachable, zero recursive returns, unchanged API-entry ESP and all five prospective callback-frame DWORDs are required. Callback code writes probes at pre-call ESP-40 through ESP-24; SendMessage pushes its four arguments and return at ESP-20 through ESP-4. The edit wrapper's five pushes therefore target exactly those probes if it writes before failed acquisition. Unrelated early/late traps fail counts or canaries. The trapped instance is discarded; recovery is not claimed. - Error: the real callback writes a chunk and returns 37. Public result 0, dwError 37, one callback and empty document must agree; source discards the failed chunk before adding its byte count.
The acquisition/retained-identity/source-routing checks remain supplementary text checks. They do not prove WAT semantics or exercise mismatched identities. No synthetic observation is a guest result. Callback ABI pointer observations are nonzero checks plus observable buffer/count effects, not full address identity proofs. EFLAGS, identity/epoch fault injection, spill exhaustion, SEH, DLL RichEdit and Windows parity remain outside scope. No claim of universal false-pass immunity is made.
Checks and retained evidence
scratch/runs/20261023-richedit-fixture-independent-review/ contains the
independent reviewer review.js, independent-checks.json, command logs,
checks.json, measured allocation, and original/corrected offline receipts. The original suite replay passed 59 controls; the corrected suite passed 70. The independent review passed 224 rejecting controls across all scenarios:
each restored outer register/control field, nested snapshots and outcomes,
all five capacity slots, callback counts/results/errors/text/ABI, slice/spin
substitutions, and every sibling result/state requirement. The actual pre-fix
baseline rejects; its SHA256 remains
420ae642fef0b7b781188875be698ca3cafeb779191f0da36ffa3baca02091e8. Syntax and diff checks pass. Missing module provenance exits 1 before VM setup.
The 56 regenerated binary artifacts (two replays of seven times four files)
matched accepted donor artifacts. Replayed binary/disassembly packages were
removed after inspection; reuse the exact sealed donor package at
/home/user/wine-assembly-rich-edit-native-callback-fixture-oct23/scratch/runs/20261023-richedit-stream-native-fixture/generated/. This review does not duplicate its packaging or change its accepted receipts.
Next exact native observations requirement
Separately authorize a runtime allocation with the normal host dependency
closure and 512 MiB guest memory. Supply a preserved normal full-module build
whose complete source closure matches the selected source revision; verify
the external build receipt, source identities, region layout and module SHA256. Do not compile or substitute historical ICM/Shadow Man modules for this review. Run the corrected fixture with WINE_ASSEMBLY_WASM,
WA_RICHEDIT_WASM_SHA256, WA_RICHEDIT_SOURCE_REV and a fresh --artifacts
directory, as specified in the donor note. Require all eight child cases
(seven scenarios and fresh ordinary after terminal capacity), all per-call
sibling assertions, watchdog completion, inspected raw observations and an
overall native PASS. Retain and diagnose the first concrete failure.