Shipwright/torch-migration/PHASE0.md

Phase 0 — Verification gates

Detail for PLAN.md Phase 0. Nothing in Shipwright changes until every gate here is green. All work happens in briaguya0/zapd-to-torch-test-harness.

---

Why there are gates at all

The harness proves 14/14 byte-for-byte parity — but for exactly one build of Torch, driven exactly one way:

	Harness measured	Shipwright will use
Build kind	executable (USE_STANDALONE=ON)	static lib (OFF)
Games compiled in	all 9 (defaults)	OoT only
Build type	Debug (-g)	Release (-O3)
Driver	main.cpp + CLI11	our own Companion calls
Extractions per process	one	up to two (vanilla, then MQ)
soh.o2r	n/a (OTRExporter makes it)	a new in-tree packer

Every row is an unproven variable. Each gate closes exactly one of them, one at a time, so a failure names its own cause. Running only the final combination would tell us "something broke" without saying what.

Reading the source says all six should be fine. Reading is how we got here; running is how we know.

---

Preconditions

• Everything builds inside the soh distrobox (distrobox enter soh) — cmake, ninja, and the compilers are not on the host PATH.
• Harness at ~/code/zapd-to-torch-test-harness, submodules initialised.
• roms/ populated: 19 ROM dumps → 14 version directories (ntsc_1-0, ntsc_1-1 and ntsc_1-2 have two dumps each; pal_mq_dbg has three). (PLAN.md says 17 hashes in a couple of places — it's 19. Corrected there.)
• assets/yml/ generated: 20,353 .yml / 119 MB across the 14 version dirs, plus the committed config.yml. Gitignored — regenerate with zapd_to_torch.py if absent.
• o2r/ holds all 19 OTRExporter reference archives (one per ROM dump, from Shipwright 95d8f7e).
• Existing baseline build at torch/build/ (USE_STANDALONE=ON, all games ON, Debug, PORT_VERSION_ENDIANNESS=ON) — leave it alone; it is the control.

---

How the gates are actually run

tools/test_assets.py, unfiltered — the same tool that produced the existing 14/14. Using anything else would make the gate results not directly comparable to the baseline they're being measured against.

An unfiltered run is a genuine full-tree, both-directions comparison:

• list_assets with no --category/--file/--type returns every key in manifests/<version>.json — all 35,386 for pal_gc.
• collect_yaml_files reverse-maps those asset paths to yml files (1,320), then setup_scratch's copy_yaml_with_externals recursion follows external_files: and pulls in the remaining 130 — all 1,450 land in the scratch dir. Verified by replaying the logic: copied == on-disk, zero yml files omitted.
• It hashes what Torch produced and diffs it against the manifest in both directions — not generated (in the reference, missing from ours) and not in reference (extras, computed over the whole zip, not just the filtered set).
• The destdir is a fresh mkdtemp per run, which matters: Process() writes destdir/torch.hash.yml and reads it back next run to skip unchanged files. Reuse a destdir and the second run silently produces a partial archive. (Same constraint that forces Phase 3's Mkdtemp().)

The Copying 1320 YAMLs to scratch dir line in the logs is printed before the external_files recursion runs — it's the pre-recursion count, not what actually gets copied. Easy to misread as a filtered subset. It isn't one.

One harness patch is needed: the torch binary path is hardcoded in two places — tools/test_assets.py:34 and lib.sh:12. Make both honour a TORCH_BIN environment override so Gate A's driver can be dropped in without editing files between runs:

TORCH = os.environ.get("TORCH_BIN") or os.path.join(SOH_DIR, "torch", "build", "torch")

./check.sh stays as the secondary whole-archive check — it compares two extracted .o2r directly, so it's the tool to reach for if a gate fails and you want the raw file-list diff rather than a per-asset table.

New: tools/matrix.sh

There's no 19-ROM loop script today; every gate needs one. It's a thin wrapper, not new comparison logic:

TORCH_BIN=<binary> tools/matrix.sh <label>

• loops all 19 ROM dumps, mapping each to its --rom-version
• calls test_assets.py <rom> --rom-version <ver> --failures-only
• writes logs/matrix-<label>/<rom>.log, prints a 19-row PASS/FAIL table, exits non-zero on any failure

Cost: ~20 s per ROM (19 s of that is Torch), so ~7 min per configuration. Four configurations plus preflight is well under an hour of wall-clock; the builds dominate.

---

Preflight — reproduce the known-good result

Before changing a single variable, re-run the current configuration on this machine, today:

TORCH_BIN=torch/build/torch tools/matrix.sh baseline

Expect 19/19 0 failed, 0 not generated, 0 not in reference. Without this, a later gate failure is ambiguous between "the variable broke it" and "something drifted in the environment."

Also confirm the reference archives are what we think they are (the README claims this for pal_gc; spot-check one more):

./manifest.sh o2r/ntsc_u_gc_b82710.o2r /tmp/regen.json
diff <(jq -S . manifests/ntsc_u_gc.json) <(jq -S . /tmp/regen.json) && echo MATCH

---

Gate D — the upstream pin (already green)

HarbourMasters/Torch main and the harness's pinned fork commit 9422bf4 have identical trees (5d247f78…; git diff --stat = 0 lines). Nothing needs upstreaming for parity.

Action: record the merge SHA 4cae44160693e1beb562e39dc301bd42278be1f9 and re-run git diff 4cae4416 9422bf4 immediately before opening the Shipwright PR, in case main has moved. No build, no matrix run.

---

Gate A — static lib, driven the way SoH will drive it

Variable: USE_STANDALONE=OFF and the hand-rolled driver, together.

They're one gate on purpose. Testing "does the static lib link" without also testing "does driving Companion by hand produce the same bytes" would leave the more interesting half unproven — and the driver has to exist either way, because USE_STANDALONE=OFF compiles out int main.

The driver is a dry run of Phase 3

torchlib-driver/main.cpp is not throwaway scaffolding. It executes the exact sequence soh/soh/Extractor/TorchExtract.cpp will use, so if any assumption in Phase 3 is wrong, it fails here — offline, against a reference archive — instead of in the game with a progress bar spinning.

zapd-to-torch-test-harness/
  torchlib-driver/
    CMakeLists.txt
    main.cpp

CMakeLists.txt — note it deliberately does not set the BUILD_<game> flags or the build type; those are Gate B's and Gate C's variables and get passed on the command line:

cmake_minimum_required(VERSION 3.26)
project(torchlib-driver CXX)
set(CMAKE_CXX_STANDARD 20)

set(USE_STANDALONE          OFF CACHE BOOL "" FORCE)
set(PORT_VERSION_ENDIANNESS ON  CACHE BOOL "" FORCE)
set(ROM_CRC_BSWAP           OFF CACHE BOOL "" FORCE)
set(BUILD_UI                OFF CACHE BOOL "" FORCE)
set(BUILD_STORMLIB          OFF CACHE BOOL "" FORCE)

add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/../torch ${CMAKE_BINARY_DIR}/torch)
add_executable(torchlib-driver main.cpp)
target_link_libraries(torchlib-driver PRIVATE torch)

torch exports src/, lib/, and the yaml-cpp includes as PUBLIC under USE_STANDALONE=OFF (torch/CMakeLists.txt:437-442), and spdlog becomes PUBLIC too — so nothing else is needed.

main.cpp — must be argv-compatible with torch o2r -s S -d D -u V <rom>, because that's exactly what test_assets.py:run_torch execs ([TORCH, "o2r", "-s", …, "-d", …, "-u", "9.2.3", rom]). Get that right and TORCH_BIN is the only thing that changes between the baseline and every gate:

#include <atomic>
#include <cstdio>
#include <filesystem>
#include <string>
#include "Companion.h"

static int RunOnce(const std::string& rom, const std::string& src,
                   const std::string& dest, const std::string& ver) {
    std::atomic<size_t> phases{ 0 };
    try {
        auto* c = new Companion(std::filesystem::path(rom), ArchiveType::O2R, false, src, dest);
        Companion::Instance = c;              // bare global; factories deref it unconditionally
        c->SetVersion(ver);
        c->SetPhaseCallback([&phases](int) { ++phases; });
        c->Init(ExportType::Binary);          // Init is the whole run; it calls Process() itself
    } catch (const std::exception& e) {
        fprintf(stderr, "driver: exception: %s\n", e.what());
        return 2;
    }
    if (!std::filesystem::exists(std::filesystem::path(dest) / "oot.o2r")) {
        fprintf(stderr, "driver: no oot.o2r produced\n");   // Process() returns void and has
        return 3;                                           // several log-and-return failure paths
    }
    fprintf(stderr, "driver: phases=%zu\n", phases.load());
    return 0;
}

…plus ~20 lines of argv parsing and, for Gate A2, a --second <rom> <dest> flag that calls RunOnce a second time in the same process.

Four Phase 3 claims this proves or refutes:

1. Companion::Instance can be assigned by hand. It's a raw global defined at src/main.cpp:15, outside the #if defined(STANDALONE) guard — so the static lib carries the storage but no entry point. Deliberate, but untested from outside.
2. try/catch is sufficient. 127 throw std::runtime_error sites across torch/src and nothing catches at the top level; an escape is std::terminate, not a popup.
3. Statting the output is necessary. Process() returns void and just logs-and-returns on missing config.yml (Companion.cpp:1304), unknown ROM hash (:1334), no config: node (:1390), bad GBI (:1484). A silent no-archive must be caught.
4. SetPhaseCallback is a usable progress denominator. One call site (gPhaseCallback(2), Companion.cpp:1282-1283), fired once per yml file at the parse→export transition. Compare the printed phases= against find assets/yml/<version> -name '*.yml' | wc -l (1,449–1,480 depending on version). If they match, the Phase 3 progress bar is monotonic 0→100; if they don't, the denominator theory is wrong and it's better to know now.

What Gate A is hunting

libgfxd. This is the one that could actually fail. torch/CMakeLists.txt:60-81 fetches and compiles libgfxd only if(USE_STANDALONE) — under OFF it isn't in the build at all. Reading the source says that's safe: every gfxd_* call sits inside #ifdef STANDALONE (DisplayListFactory.cpp:81-199 = DListCodeExporter + DebugDisplayList; DisplayListOverrides.cpp:18-224 = the GFXDOverride callbacks), while DListBinaryExporter (:239) and DListFactory::parse (:526) are outside it. Display lists are the largest asset class in the archive, so if that reading is wrong the gate fails immediately and unmistakably.

Other STANDALONE deltas, all expected inert:

• Companion.cpp:1524 gates enums: parsing — config.yml has no enums: key.
• DListCodeExporter isn't registered — we export Binary.
• yaml-cpp gains YAML_CPP_STATIC_DEFINE; spdlog moves PRIVATE→PUBLIC; the standalone-only find_package(ZLIB REQUIRED) disappears (zlibstatic from FetchContent remains).

One landmine worth writing down: if(NOT USE_STANDALONE AND EXISTS "/mnt/c/WINDOWS/system32/wsl.exe") (torch/CMakeLists.txt:295) FetchContents Microsoft GSL. It fires only inside WSL, and only in the static-lib configuration — i.e. exactly what Shipwright will use. Irrelevant in the distrobox; it will bite the first person who builds SoH under WSL after this lands.

Running it

cmake -S torchlib-driver -B torchlib-driver/build -GNinja -DCMAKE_BUILD_TYPE=Debug
cmake --build torchlib-driver/build -j
TORCH_BIN=torchlib-driver/build/torchlib-driver tools/matrix.sh static-lib

Pass: 19/19 identical, and phases= matches the yml count for each version. Smoke-test pal_gc alone before committing to the full matrix.

---

Gate A2 — two extractions in one process

Variable: re-entrancy. Free to add once the driver exists, and it retires PLAN.md risk #6 offline instead of by hand in the game.

RunExtract can extract more than once per process (PS_FIRST → PS_SECOND, and ES_EXTRACT_ARGS loops over CLI args), so a real user extracting vanilla and then MQ in one session hits state SoH has never exercised:

• gProcessedFiles (Companion.h:300) is never cleared — a second Process() on the same instance skips every file. Phase 3's answer is a fresh Companion per call; this gate is what proves that's enough.
• gAddrMap / gParseResults / gWriteMap also accumulate.
• AliasManager::Instance isn't cleared at the end of Process (believed inert for OoT — no OoT factory calls Register — but unproven).
• AudioManager::Instance leaks per run (Companion.cpp:1584).
• File-scope statics in TextureFactory.cpp:14-15, CompressedTextureFactory.cpp:17-18, DisplayListFactory.cpp:82.

This is the one gate test_assets.py can't drive — it execs the binary once per ROM, so two extractions in one process has to come from the driver itself. Hence the --second flag, and a --pair mode in matrix.sh that invokes the driver directly and then hands each of the two outputs to check.sh:

DEST_A=$(mktemp -d); DEST_B=$(mktemp -d)
"$TORCH_BIN" o2r -s assets/yml -d "$DEST_A" -u 9.2.3 roms/pal_gc_0227d7.z64 \
             --second roms/pal_mq_f46239.z64 "$DEST_B"
# then, for each of (DEST_A, pal_gc_0227d7) and (DEST_B, pal_mq_f46239):
cp "$DEST/oot.o2r" o2r/torch.o2r; cp "o2r/<rom>.o2r" o2r/reference.o2r; ./check.sh

One process, two fresh Companions, two destdirs. Deliberately a vanilla ROM then an MQ one — the same sequence a user hits when they answer "Yes" to "extract another". Pass: check.sh prints PASS: All N files match! for both — specifically for the second, which is the whole point.

---

Gate B — OoT-only build

Variable: the eight non-OoT BUILD_<game> flags. Still USE_STANDALONE=ON, still Debug, so this is measured against the preflight baseline with exactly one thing changed.

cmake -S torch -B torch/build-oot -GNinja -DCMAKE_BUILD_TYPE=Debug \
      -DPORT_VERSION_ENDIANNESS=ON \
      -DBUILD_SM64=OFF -DBUILD_MK64=OFF -DBUILD_SF64=OFF -DBUILD_PM64=OFF \
      -DBUILD_FZERO=OFF -DBUILD_BK64=OFF -DBUILD_MARIO_ARTIST=OFF -DBUILD_NAUDIO=OFF
cmake --build torch/build-oot -j
TORCH_BIN=torch/build-oot/torch tools/matrix.sh oot-only

Each BUILD_X=OFF drops -DX_SUPPORT and filters src/factories/x/* out of the source glob (torch/CMakeLists.txt:158-207). Two effects are real:

• BUILD_BK64=OFF removes BK64::TrySynthesizeRomConfig, Companion's fallback for a ROM hash absent from config.yml. All 19 of ours are present, so it's unreachable here — but it means an unknown dump behaves differently, which is precisely why Phase 3 must stat the output rather than trust the run.
• BUILD_NAUDIO=OFF removes Torch's only tinyxml2 consumers. It does not remove the tinyxml2 FetchContent — that's unconditional — so Phase 2's OVERRIDE_FIND_PACKAGE concern stands regardless.

Nothing else should touch a shared path. A cross-game factory registration or a shared header behind an #ifdef would show up as a diff, and that's what the gate is for.

---

Gate C — Release build

Variable: -DCMAKE_BUILD_TYPE=Release. Otherwise the baseline configuration.

This gate isn't in PLAN.md and should be. The harness's 14/14 is a Debug measurement (torch/build/CMakeCache.txt: CMAKE_BUILD_TYPE=Debug), and so is the local Shipwright build. Shipwright's CI and every release build Release — Torch sets CMAKE_CXX_FLAGS_RELEASE "-O3" for non-MSVC (torch/CMakeLists.txt:242). Any UB, uninitialised read, or unspecified evaluation order in Torch can change bytes between -g and -O3.

cmake -S torch -B torch/build-release -GNinja -DCMAKE_BUILD_TYPE=Release -DPORT_VERSION_ENDIANNESS=ON
cmake --build torch/build-release -j
TORCH_BIN=torch/build-release/torch tools/matrix.sh release

Seven minutes to find out, versus finding out from a user whose release-build archive doesn't match their friend's. If it fails it's an upstream Torch bug and everything else stops until it's fixed.

---

Gate A′ — the shipping configuration

Variable: all of them at once — static lib + driver + OoT-only + Release.

Not implied by A ∧ A2 ∧ B ∧ C. #ifdef interactions are combinatorial, and this is the only configuration that will ever ship.

cmake -S torchlib-driver -B torchlib-driver/build-ship -GNinja -DCMAKE_BUILD_TYPE=Release \
      -DBUILD_SM64=OFF -DBUILD_MK64=OFF -DBUILD_SF64=OFF -DBUILD_PM64=OFF \
      -DBUILD_FZERO=OFF -DBUILD_BK64=OFF -DBUILD_MARIO_ARTIST=OFF -DBUILD_NAUDIO=OFF
cmake --build torchlib-driver/build-ship -j
export TORCH_BIN=torchlib-driver/build-ship/torchlib-driver
tools/matrix.sh shipping
tools/matrix.sh --pair pal_gc_0227d7 pal_mq_f46239 shipping-pair   # Gate A2, shipping config

When A′ is green, that flag set is the spec. Phase 2's root-CMakeLists.txt block is a transcription of it, not a re-derivation — and main.cpp's RunOnce is a transcription into TorchExtract.cpp. Copy, don't rewrite.

---

Gate E — the soh.o2r baseline

Different in kind from the others: it doesn't test Torch at all. It captures the artifact the Phase 4 packer has to reproduce, from a pipeline that is about to be deleted. Do this while ZAPD still builds.

Current state: soh.o2r at the repo root, 1,042 entries = 1,041 files under soh/assets/custom/ + portVersion. No version file — matching sohArchiveVersionMatch (OTRGlobals.cpp:282-283), which only reads portVersion.

1. Rebuild rather than trust the file on disk, so provenance is unambiguous:

   cmake --build build --target GenerateSohOtr     # inside the soh distrobox
   

   GenerateSohOtr writes into the source tree: it rm -rfs and re-copies soh/assets/custom/shaders/ from libultraship/src/fast/shaders/ (3 files — directx/default.shader.hlsl, metal/default.shader.metal, opengl/default.shader.glsl). That directory is gitignored (soh/assets/.gitignore:8), which is why git status stays clean and why the on-disk count (1,041) exceeds git ls-files (1,038). Phase 4's packer must run after the same copy or the archives differ by exactly those 3 entries.

2. Record the output — manifest.sh already does exactly this job:

   ./manifest.sh /var/home/briaguya/code/Shipwright/soh.o2r manifests/soh_o2r.json
   

   Lives in the harness alongside the other 14 manifests. (Not in the yml repo — that holds yml and nothing else.)

3. Record the inputs too, or a future mismatch is ambiguous between "the packer is wrong" and "the assets changed":

   • sha256 of every file under soh/assets/custom/ → manifests/soh_custom_inputs.json
   • CMAKE_PROJECT_VERSION = 9.2.3 (CMakeLists.txt:9), the portVersion payload
   • the develop SHA the archive was built from

4. Call out the 31 files that matter. Only these hit the conversion path — everything else is a byte copy, and a byte copy that fails is a trivial bug:

   Format	Count
   .rgba32.png	15
   .ia8.png	8
   .ia16.png	5
   .ia4.png	3

   The other 1,010: 923 objects, 38 plain .png, 23 .rgba16, 22 json, 4 ttf, 3 shaders, 2 xml. The trap is quantisation, not decoding — n64graphics' SCALE_8_4(x) (x*15/255) vs ZAPD's (x >> 4) & 0xF disagree (0xF0 → 0xE vs 0xF). Phase 4 hand-ports ZAPD's arithmetic; this manifest is what catches it if the port is wrong.

---

Exit criteria

Gate	Pass condition	Status
Preflight	19/19 identical with the existing baseline binary; one manifest re-derivation matches	✅ green — see below
D	pin SHA recorded; git diff vs upstream main still empty	✅ green (trees equal)
A	19/19 identical; phases= matches the per-version yml count	✅ green — see below
A2	both archives in a two-extraction process identical — especially the second	✅ green — see below
B	19/19 identical	⬜
C	19/19 identical	⬜
A′	19/19 identical, plus the pair run	⬜
E	manifests/soh_o2r.json (1,042 entries) + input hashes + version + SHA committed	⬜

"Identical" means test_assets.py reports 0 failed, 0 not generated, 0 not in reference — not "close enough". --failures-only keeps the logs to the point; drop it when bisecting.

Write the results up as a second table in the harness README, next to the existing 14/14 one. That table is the evidence the Shipwright PR points at.

Preflight result — 2026-07-24

TORCH_BIN=torch/build/torch tools/matrix.sh baseline, harness phase0-gates off bump-torch-9422bf4. 19 passed, 0 failed, of 19 ROM dumps, every one 0 failed, 0 not generated, 0 not in reference.

Version	Dumps	Assets
ntsc_1-0	ntsc_j_1-0_c892bb, ntsc_u_1-0_ad69c9	38,390
ntsc_1-1	ntsc_j_1-1_dbfc81, ntsc_u_1-1_d3ecb2	38,526
ntsc_1-2	ntsc_j_1-2_fa5f59, ntsc_u_1-2_41b3bd	39,032
ntsc_j_gc	ntsc_j_gc_0769c8	39,064
ntsc_j_gc_collection	ntsc_j_gc_collection_2ce2d1	39,066
ntsc_j_mq	ntsc_j_mq_dd14e1	39,026
ntsc_u_gc	ntsc_u_gc_b82710	39,066
ntsc_u_mq	ntsc_u_mq_8b5d13	39,029
pal_1-0	pal_1-0_328a1f	35,362
pal_1-1	pal_1-1_cfbb98	35,362
pal_gc	pal_gc_0227d7	35,386
pal_gc_dbg	pal_gc_dbg_cee6bc	35,646
pal_mq	pal_mq_f46239	35,352
pal_mq_dbg	pal_mq_dbg_079b85, _50bebe, _cfecfd	35,612

Counts match the harness README's 14-target table exactly. The 19-dump matrix adds something the 14-target one never showed: the duplicate dumps agree with each other — both 1.0 dumps produce the same 38,390 assets against the same manifest, and all three pal_mq_dbg dumps produce the same 35,612. Since PLAN.md's MQ handling keys off IsMasterQuest() rather than the dump, that's worth having on record.

Reference archives spot-checked as well — manifest.sh re-derivation of o2r/ntsc_u_gc_b82710.o2r (39,066 entries) and o2r/pal_mq_dbg_079b85.o2r (35,612) reproduces the committed manifests byte-for-byte. The baseline the gates measure against hasn't drifted.

Gate A result — 2026-07-24

TORCH_BIN=torchlib-driver/build/torchlib-driver tools/matrix.sh static-lib, harness phase0-gates @ 504f086. 19 passed, 0 failed, of 19 ROM dumps — identical counts to the preflight table above, every one 0 failed, 0 not generated, 0 not in reference.

USE_STANDALONE=OFF changes nothing about the bytes, and driving Companion by hand produces exactly what the CLI does. Confirmed rather than assumed:

• libgfxd really is absent. torchlib-driver/build/_deps/ contains only tinyxml2, yaml-cpp and zlib — no libgfxd-src, because the FetchContent sits inside if(USE_STANDALONE). The Binary export path is genuinely gfxd-free; all 19 archives extract identically without it.
• The progress denominator is exact. phases= matched find assets/yml/<version> -name '*.yml' | wc -l for all three distinct counts: pal_gc 1450/1450, ntsc_u_gc 1449/1449, pal_gc_dbg 1480/1480. So Phase 3's bar is monotonic 0→100 with a denominator counted off disk, and OTRGlobals.cpp:750-755 needs no change.
• Archive size matches the CLI build's (33,154,599 bytes for pal_gc). It's ~343 KB larger than the OTRExporter reference: 76 KB of that is Torch's duplicate archive entries (see Gate A2 below) and the remaining ~267 KB is zip-compression difference, not content. Which is why comparison is always file-by-file inside the archive.
• The Companion::Instance hand-assignment, SetVersion, SetPhaseCallback, Init(Binary) sequence works from outside the library, so TorchExtract.cpp can be a transcription of RunOnce().

Also learned, for Phase 2: zlib is fetched under USE_STANDALONE=OFF with BUILD_STORMLIB=OFF — _deps/zlib-src exists despite zero zlib.h includes in torch/src, because that FetchContent_Declare(... OVERRIDE_FIND_PACKAGE) is unconditional. Risk #1 is observed, not hypothetical. Conversely spdlog was not fetched (find_package(spdlog QUIET) found the distrobox's), so the second-copy ABI concern doesn't apply — only the global-logger stomping.

Gate A2 result — 2026-07-24 — ✅ green

TORCH_BIN=torchlib-driver/build/torchlib-driver tools/matrix.sh --pair pal_gc_0227d7 pal_mq_f46239 reentrancy. One process, vanilla then MQ, a fresh Companion and destdir each:

	assets	missing	extra	mismatched
pal_gc (first)	35,386 / 35,386	0	0	0
pal_mq (second)	35,352 / 35,352	0	0	0

The load-bearing number is the second run's phases=1450, exactly pal_mq's on-disk yml count. Had gProcessedFiles leaked across runs the second pass would have skipped files it believed already processed — a collapsed phase count and a truncated archive. It processed all 1,450 and reproduced the reference exactly.

A fresh Companion per extraction is sufficient. AliasManager::Instance not being cleared, AudioManager::Instance leaking, and the file-scope statics in TextureFactory.cpp:14-15, CompressedTextureFactory.cpp:17-18, DisplayListFactory.cpp:82 are all confirmed inert for OoT/Binary. PLAN.md risk #6 is retired offline; the in-game two-ROM session (§9 of the verification matrix) is now a confirmation, not a discovery.

Found along the way: Torch writes duplicate archive entries

The first attempt failed both comparisons — not on parity, but because check.sh never got to compare. unzip had aborted mid-extract.

Torch writes some assets more than once under the same archive path. pal_mq: 35,377 physical zip entries for 35,352 unique names; 24 names duplicated (23 cutscenes + link_home_room_0Back ground_002480, which appears three times) = 25 redundant entries. pal_gc: also 24. Every copy is byte-identical to its siblings. All 19 OTRExporter reference archives have zero. Identical under USE_STANDALONE ON and OFF, so it is not a static-lib artifact — pre-existing Torch behavior, and Gate A's conclusions are unaffected.

Zip permits this: the format is a flat sequence of member records plus a central directory, with the path as an ordinary string field and no uniqueness constraint. Only a reader materializing to a real filesystem has to reconcile them — hence unzip prompting replace ...?, reading EOF from a non-interactive stdin, and exiting non-zero into set -e.

Benign for SoH. O2rArchive::LoadFile resolves via zip_name_locate, which returns one entry index, and the copies are byte-identical. O2rArchive::Open is the only place LUS walks entries by raw index, and it only feeds Archive::IndexFile, which does (*mHashes)[CRC64(path)] = path — a repeat insert is an idempotent overwrite. ListFiles/HasFile derive from that same map. So the duplicates collapse at the archive's front door: no double-loading, no double-counting, no duplicate list entries. Cost is 76 KB compressed on a 33 MB archive.

Filed as item 5 of HarbourMasters/Torch#233 for post-migration cleanup. Not a blocker and not something the migration needs to fix.

The lasting lesson is about the harness, not Torch: a comparison keyed on archive path — which is both test_assets.py and check.sh — structurally cannot see duplicate entries, because they collapse into one key. Asset parity genuinely holds (identical unique-path sets, identical bytes at every path), but "19/19 identical" was never a statement about the zip container. check.sh now extracts with unzip -o and prints a duplicate-name count for both archives so this can't go quiet again.

---

Cost

	Build	Matrix
Preflight	—	~7 min
A	full Torch rebuild	~7 min
A2	— (reuses A)	~1 min
B	full Torch rebuild	~7 min
C	full Torch rebuild	~7 min
A′	full Torch rebuild	~8 min
E	Shipwright GenerateSohOtr	~1 min

An evening, dominated by four full Torch builds. Four extra build trees; the Debug Torch binary alone is 125 MB, so either build them sequentially and delete as you go or budget a few GB.

---

What Phase 0 deliberately does not cover

Being explicit so these don't get mistaken for retired risk:

• Linking Torch into Shipwright. The zlib / stb / StringHelper / StrHash64 / tinyxml2 / storm collisions in PLAN.md Phase 2 need SoH's own build to surface. A standalone driver links Torch against nothing else, so it can't see any of them. That's Phase 2's first day.
• spdlog global-state stomping. Torch calls spdlog::set_level/set_pattern on the default logger (Companion.cpp:193-194, 1810-1814), and our config.yml sets logging: CRITICAL — so this fires on every run, it just doesn't matter until there's a SoH logger to clobber.
• Windows and macOS. Every gate runs Linux/gcc in the distrobox. Torch forces /MT under MSVC, adds -DSTORMLIB_NO_AUTO_LINK and /bigobj, and carries a cmake_minimum_required(3.12) + CMAKE_POLICY_VERSION_MINIMUM shim under a CMake 4 host. If a Windows machine is available, running Gate A′ there is the cheapest possible insurance on the one platform that can't be predicted from reading — but it isn't a blocker for starting Phase 2.
• Anything in-game. The driver proves the bytes; it says nothing about the ImGui progress bar, the ROM picker, or ValidHashes. That's the Phase 3 manual matrix.