Pawn fourth record: native writer map; runtime trace capacity refused
The fourth record and second reply remain unresolved. No emulator defect was
demonstrated and no production change was made. Evidence is in ignored
scratch/runs/20261113-pawn-fourth-writer-1520/.
Read parent 5ddfd2b93, verified all 169 parent SHA seals, and independently
hashed its complete authentic 100-byte save to
811659988ee98f19ca08615f1191431993f5d51646ea41efed5fdb6f0ff5cf77.
The original file remains at
scratch/runs/20261113-pawn-modal-guard-1445/pawn-1445-secondmove.gam;
it was neither edited nor copied. Its first three records authenticate
e2-e4, g8-f6 and d2-d4. Record four is still exactly 0,44,7,15,0,0.
No third move was sent.
What the bounded native analysis establishes
Disassembly and literal cross-references use the original Pawn.exe with SHA
21f78408e306d77103484bc90accfd511a141131ed1e14990380ecea19de5a58.
Small gzip listings preserve bounded instruction ranges and commands are
reproducible with node tools/disasm_fn.js EXE VA COUNT --linear and
node tools/xrefs.js EXE VA --code. Cross-reference classifications are
heuristics; actual instruction listings support the statements below.
The magic-77 save loop at 0x40e6ce serializes six parallel history arrays
through globals 0x4231b0/b8/c0/c8/d0/d8. It increments a local index from
zero to one, writes all six DWORDs, then checks the next origin for -1.
It does not use the current history cursor 0x422fa8, validate endpoints,
or test a separate complete-record flag. Thus the authentic file proves
stored array contents, not by itself successful move execution. Neither
zero origin nor the other fields is interpreted specially by this writer.
This does not exclude special semantics elsewhere in the guest.
The native AI commit path is now located:
| Address | Native operation |
|---|---|
0x40299c |
Increment history cursor 0x422fa8. |
0x4029e2 |
Call 0x4038fa with selected endpoints (display/animation path). |
0x402a1d |
Call 0x408ef7 with the same selected endpoints. |
0x402a4c |
Write origin from signed byte [DWORD(0x4230f8)+DWORD(0x4230fc)+1] into current history origin. |
0x402a7a |
Write destination from signed byte [DWORD(0x423104)+DWORD(0x423108)+1] into current history destination. |
0x402a96..0x402aea |
Copy four move-state values from 0x4230c8/0x423148/0x423138/0x423140 pointees into the remaining history arrays. |
0x402b19 |
Write next history origin -1. |
0x402b3c/0x402b5d |
Copy the same selected endpoints to globals 0x42302c/0x423030, followed by display update 0x404dec. |
These are identified writer sites, not an observed first write to record
four. The player path has distinct stores at 0x40df77/0x40df8f; file
loading and other native paths also access these arrays. No claim that the
AI sites actually produced this particular record follows from static
references alone.
The selected-endpoint arrays are allocated with stride 101 at
0x4011d2/0x40120f. Search writes their principal variation at 0x41089b;
0x4108c3/0x4108e9 write endpoint bytes using stride times search ply plus
ply, and subsequent code copies continuation bytes. AI entry clears the
root origin row at 0x402447..0x402495; that loop does not clear the
destination row. This supplies a concrete reason to inspect the selected
origin and destination separately, but does not establish that a cleared
origin caused record four.
AI selection first tries opening-book routine 0x40b467 when history
cursor is below 40, 0x423008 is zero and 0x42301c is nonzero. A nonzero
return sets 0x423078=1 and jumps to commit. Otherwise 0x4025da calls
search 0x40ea8c; its score goes to 0x42307c. Search iteration exit tests
depth 0x42306c against 0x422f7c, elapsed time against 0x422f78, score
magnitude and 0x423038. Search itself also checks abort/time at
0x40eaa1 and 0x40eb35. No clock or search setting was changed.
The zero-delimited original opening book is read bytewise by native code
at 0x4072ff, and 0x40b50f..0x40b573 compares endpoint bytes to history.
Offline prefix comparison finds 1,155 entries beginning e4 and 41 beginning
e4/Nf6, but zero beginning e4/Nf6/d4. See
offline-book-prefix.json. Conditional on an intact runtime book and the
authenticated history, this points to the search fallback for the second
reply. The actual runtime book buffer, branch and selected endpoints were
not observed, so this remains a conditional deduction.
Native make-move code 0x4083d8 saves move state and has explicit en-passant,
promotion and castling-related paths. The notation routine 0x408ef7
ultimately calls it at 0x409357, after resetting ply to zero. At
0x408ccf..0x408cef, make-move saves the destination's previous piece code
through 0x4230c8; at 0x408cf3..0x408d0e it transfers the origin piece to
the destination and writes empty code 7 to the origin. Together with the
history copy, this identifies the third serialized DWORD as the saved
destination piece code, rather than a record-complete flag. The fourth
comes from the saved 0x422f88 state (0x408406..0x40841c), whose bits are
cleared for king/rook movement in the castling paths. The fifth tracks the
double-pawn destination (0x4088fc..0x408910); the sixth is promotion state
(0x408e44..0x408e53). These explain ordinary state fields but do not prove
why this record has origin zero. Its tail 0x408ed7..0x408ee6
increments search ply and flips side 0x422f5c as 1-side; unmake
0x403117..0x403126 reverses ply and side. The outer AI loop at
0x40232b calls the responsive input loop 0x40ca11, with idle Sleep(10)
at 0x40ca35. Reaching that idle loop alone cannot prove a valid second
reply or identify the earlier instruction/API that selected an endpoint.
Prepared observer and concrete blocker
Reused the qualified 1445 owner guard, finite drags with independent UP,
strict modal guards and no-redundant-motion helper unchanged. No previously
qualified cases were rerun. The additional read-only history-read.js
records history, side, cursor, search depth/score, book flag and selected
endpoint bytes. Existing interpreter --trace-at=0x402a50 was prepared to
capture up to four AI commit hits immediately after the origin store, with
indirect history/selected-array dumps and relevant globals. It would
distinguish the source byte from the stored origin without changing guest
memory, position, queues, AI configuration or clock. It was not run.
The freeze preserves Git 5ddfd2b938b291764767bb9b785461f485c0e747, relevant
binary diff, helper patches and 321 actual source/module/media pins. Module
SHA remains 6888cc5e4ec7693a8a317fe29147994c7f88fe43afe3109fe6ab0e3e349eec9f.
Concurrent WebGL changes were separately owned and recorded read-only;
all 321 frozen contents still matched after their publication. No source
archive or repeated media was created.
After LMarbles released the heavy slot, the supervisor acquired
/tmp/wine-heavy-runtime.lock, verified zero actual test/run.js children,
then refused its sole launch decision at 1,499,303,936 available bytes,
696,064 bytes below the required 1,500,000,000 floor. Neither Pawn nor its
watchdog was created, and no input occurred. Two subsequent read-only
control requests received connection-refused because no server existed;
they did not launch or replay anything. No floor retry, foreign cleanup,
box, substitute DLL or guest patch followed.
Supervisor PID 1108660 actually waited 0 with separate stdout/stderr EOFs; outer tool session 66475 completed exit 0. Native/watchdog waits are N/A (not started), not successful runtime receipts. Eleven helper syntax checks, 26 gzip decode/hash checks, the parent seals and frozen-content checks passed. This is documentation-only and claims no new build or runtime acceptance.
Next: sufficient sustained disk headroom, a newly frozen guarded original, reviewed e4/Nf6 then d4, and the prepared bounded interpreter trace. Observe the first actual record-four writer and selected-byte provenance before attributing a failure to any emulator instruction or API. No minimal failing regression or scoped fix is justified yet. Box ID none, state NOT_CREATED, expiry null, stop receipt N/A.