add docs on shared architecture

AGENTS.md

@@ -11,6 +11,10 @@ primary evidence.
 The server binaries are **C++** (MSVC 2010, x64). See
 [`docs/ARCHITECTURE.md`](docs/ARCHITECTURE.md) for the evidence and what is known
 about their structure.
+All servers share one in-house framework (`Magi*`): server-kind enum,
+startup modes, XML option files and DB access via DBC servers. See
+[`docs/SHARED_FRAMEWORK.md`](docs/SHARED_FRAMEWORK.md) and check each new
+binary against it instead of re-deriving it.
 
 ## Target runtime environment
 
@@ -62,19 +66,35 @@ other licence-restricted vendor files.
 
 ## Analysis workflow
 
+The full reproducible procedure, with commands, is in
+[`docs/WORKFLOW.md`](docs/WORKFLOW.md). Follow it for every binary. In summary:
+
 1. Route the target first: native executables/libraries go through REA
    `open_binary` (never pass a Ghidra project file as the target — pass the
    original binary path under `binaries/`).
 2. Work summary-first: `binary_overview` / summary detail before drilling into
    `analyze_function` or pseudocode. Reuse retained Evidence IDs instead of
    re-running identical analysis.
-3. Export raw decompilation artifacts (pseudocode, listings) into `decomp/`
+3. **Third-party libraries: every time one is discovered** (an imported DLL, a
+   version string, RTTI class prefix, library error strings, or a PDB path),
+   do the following before continuing analysis:
+   - pin its version;
+   - **search online for its source or headers** (official releases, git tags,
+     vendor repos, clean-room equivalents);
+   - download it into `third-party/`, or into the gitignored
+     `third-party/_local/` if its licence forbids redistribution;
+   - record it in `third-party/README.md`.
+
+   If no source exists, record that and the acquisition routes. See
+   `docs/WORKFLOW.md` §3.
+4. Export raw decompilation artifacts (pseudocode, listings) into `decomp/`
    using per-binary subdirectories, e.g. `decomp/GameServer_Release_USA_x64/`.
    Keep filenames stable and function-address-based so results are reproducible.
-4. Track targets, artifact SHA-256 identities, and per-binary status in
+   For full coverage, use the headless scripts in `decomp/scripts/`.
+5. Track targets, artifact SHA-256 identities, and per-binary status in
    `decomp/INVENTORY.md`; keep a per-binary finding ledger in
    `decomp/<name>/evidence.md`.
-5. Close the REA session (`close_binary`) when the investigation is complete.
+6. Close the REA session (`close_binary`) when the investigation is complete.
 
 ## Evidence and claims
 

docs/ARCHITECTURE.md

@@ -3,6 +3,9 @@
 What is known about how the shipped server binaries are built and structured.
 Each claim is marked **observed**, **inference**, or **unknown** (see AGENTS.md).
 
+Shared framework (config files, server kinds, DB access path): see
+[`SHARED_FRAMEWORK.md`](SHARED_FRAMEWORK.md).
+
 ## Implementation language: C++
 
 All 16 server executables in `binaries/` are written in **C++**, compiled with

docs/SHARED_FRAMEWORK.md

@@ -0,0 +1,251 @@
+# Shared server framework ("Magi") — configuration and database access
+
+The server EXEs share one in-house framework. Its classes are named `Magi*`,
+and the network sessions are `MG*Session`. This document defines the parts
+every binary shares:
+- server identity;
+- startup modes;
+- configuration files;
+- the path to the database.
+
+Analyse each new binary against this document, and record in its own
+`evidence.md` where it differs, rather than re-deriving the framework.
+
+Status: first pass. **Primary evidence:** `MessageServer_Release_x64.exe`
+(`decomp/MessageServer_Release_x64/`). **Cross-binary checks:** strings and RTTI
+only, unless a row says otherwise. Field offsets are from MessageServer. Other
+binaries share the same code, so they are *expected* to match, but this hasn't
+been verified. Every claim is static; nothing here was observed at runtime.
+
+Notation: **O** = observed in decompiled code, **S** = observed in
+strings/RTTI only, **I** = inference, **?** = unknown. `MS:0x…` means a
+function in `decomp/MessageServer_Release_x64/full/functions/0x….c`.
+
+---
+
+## 1. Framework presence across binaries (S)
+
+| Marker | Present in |
+|---|---|
+| RTTI `MagiServerOption`, `MagiCustomOption`, `MagiServerFile`, `MagiServerManager` | all 14 server EXEs. The two SMMonitoring EXEs only have `MagiServerManager`. |
+| Option-loader keys (`write_thread_count`, `each_option`), the `-AGNET MODE-` banner [sic], and the `DBC_CHAT`/`DBC_EMERGENCY` names | all 14 server EXEs; none in SMMonitoring |
+| `Servers2DbcProxy` RTTI (client of the DBC servers) | Auction, Dungeon, GameGateway, Game, GM, Guild, Manager, Message, PVP |
+| ADO / `_Connection` strings (direct SQL access) | DBCServer and DBCServer_LOG (4 hits each); 1–2 hits in Auction, Dungeon, Game, Guild, Manager and PVP (**?** — may be incidental) |
+
+**Inference:** only the DBC servers talk to SQL Server. Every other server
+reaches the database through ProudNet RMI to a DBC server (§5). This needs
+confirming by decompiling the ADO usage in the six non-DBC servers that have
+hits.
+
+## 2. Server-kind enum (O, `MS:0x140002930`)
+
+The one-byte server type is stored in the option object (`+0x1e8`) and in
+peer/session records. Names below are the wide strings the binary prints.
+
+| Value | Name | Value | Name |
+|---|---|---|---|
+| 1 | Lobby | 0x0c | DBC_GAME |
+| 2 | Dungeon | 0x0d | Agent |
+| 3 | PVP | 0x11 | DBC_LOG |
+| 4 | WAR | 0x12 | DBC_CHAT |
+| 5 | Guild | 0x13 | DBC_EMERGENCY |
+| 6 | Message | 0x15 | Agentmanager |
+| 7 | GM | 0x16 | MiniGame |
+| 8 | Gate | 0x17 | `api` peer (no display name; `MS:0x14005dc20`) |
+| 9 | Manager | 0x0e–0x10, 0x14 | ? |
+| 0x0a | Auction | | |
+| 0x0b | Match | | |
+
+Separately, servers also have a numeric **code** (the instance ID). It appears
+in XML as `code="61"` and in file names such as `MSG_61.xml`. In the shipped
+deployment config every kind has one code: DBC_GAME=1, DBC_LOG=2, DBC_CHAT=3,
+DBC_EMERGENCY=4, Msg=61, GM=71, Gate=81, Manager=91, Auction=101, Match=111,
+Agent=800, AgentManager=900, Dungeon=2001, PvP=3003, Guild=5001, Lobby=21001
+(S, `binaries/Option/*.xml`).
+
+## 3. Startup modes (O, `MS:0x14005c7b0`)
+
+`MagiServerOption::Init(argv1)` records the PID and the start time
+(`+0x1ec`, `+0x1f0`), then branches on the first character of the argument:
+
+| `argv[1]` | Mode | What is loaded |
+|---|---|---|
+| does not start with a digit | **standalone** | `argv[1]` is a path to a per-server XML file, which `MS:0x14005dc20` parses (§4.1) |
+| starts with a digit | **agent mode** (prints `-AGNET MODE-`) | `atoi(argv1)` is this server's **code**. `MS:0x14005ceb0` reads `option\total_option.xml` (§4.3). The full server list then arrives from the Agent over RMI: the `MGSAgentSession` vftable slot 7 is `MS:0x140065ab0`. The entry whose code equals our code configures this server (`MS:0x14005c9f0`); every other entry is added as a peer (`MS:0x14005ff50`). |
+
+This matches the Agent side: the Agent launches a server with the numeric
+server code as its argument (`decomp/Agent_Release_x64/evidence.md` finding 3).
+
+**ServerInfo record** pushed by the Agent (O, copy constructor `MS:0x140065900`; offsets in bytes):
+
+| Off | Type | Field (I) |
+|---|---|---|
+| +0x00 | int | world |
+| +0x04 | int | code |
+| +0x08 | int | ? |
+| +0x0c | u8 | server kind (§2) |
+| +0x10 | int | private_port |
+| +0x14 | int | public_port |
+| +0x18 | int | public_udpport |
+| +0x1c | int | udp_count |
+| +0x20 | u8 | ? |
+| +0x28 | `std::wstring` | private_ip |
+| +0x50 | `std::wstring` | public_ip |
+| +0x78 | `std::wstring` | name |
+
+## 4. Configuration files
+
+All of them are XML parsed with TinyXML 2.5.x (`third-party/tinyxml-2.5.3`).
+Paths are relative to the working directory, under `option\`. A sample
+deployment set ships in `binaries/Option/`.
+
+That sample set contains placeholders such as `YOUR_WAN_IP`, `DB_USER` and
+`DB_SERVER`. It is probably a community deployment, not the original publisher
+files (I). It also contains **plaintext DB passwords** — never copy them into
+docs.
+
+### 4.1 Per-server file — `MagiServerOption` (O, `MS:0x14005dc20`)
+
+Default name per binary: `option\<kind>_serveroption.xml`. In MessageServer
+this is the global `L"option\\msg_serveroption.xml"` (`MS:0x14012e5c0`).
+Per-instance copies are `<KIND>_<code>.xml`, e.g. `MSG_61.xml`.
+
+Top-level elements, the attributes read from each, and where the value is
+stored in the option object:
+
+| Element | Attribute → field |
+|---|---|
+| `server` | `name` → +0x08 wstr · `code` → +0x154 · `private_ip` → +0x30 wstr · `private_port` → +0x80 · `public_ip` → +0x58 wstr · `public_port` → +0x84 · `public_udpport` → +0x88 · `udp_count` → +0x8c · `thread_num` → +0x15c (clamped ≥1; defaults to 1 if missing) · `agent` → +0x150 (bool, `=="1"`) · `debug` → +0x165 (bool) · `world` → +0x140 · `private_dns` / `public_dns` → override +0x30 / +0x58 (I: hostname form) |
+| `db_account`, `db_web`, `db_game`, `db_game_log` | `string` → +0x90 / +0xb8 / +0xe0 / +0x108 (std::string, ADO connection string) · `timeout` → +0x130 / +0x134 / +0x138 / +0x13c (`MS:0x14005fa80`) |
+| `dbc` | `write_thread_count` → +0x14c |
+| `db_cache` | `session_count` → +0x144 (also copied to +0x148) |
+| `db_log_cache` | `session_count` → +0x148 |
+| peer elements (table below) | `ip`, `port` → peer map at +0x1a8, keyed by kind (`MS:0x14005fb60`) |
+| `dns` | `dns` → +0x151 (bool) |
+| `setting` | `multi` → +0x166 (bool) · `lock` → +0x168 (double) |
+| `auto` | `id`/`value` → +0x158 (auto-start flag) |
+| `start_server` / `revive_server` / `open_channel` | → +0x178 (double) / +0x171 (bool) / +0x170 (bool). The element-to-field pairing is from the order of assignments: **I** |
+| `send` | `name`, `minlev`, `maxlev`, `group`, `type`, `id` → ? (not yet traced) |
+
+Peer elements parsed by MessageServer, with the server kind each maps to:
+
+| Element | Kind |
+|---|---|
+| `agentmanager` | 0x15 |
+| `db_cache` | 0x0c |
+| `db_log_cache` | 0x11 |
+| `msg_server` | 6 |
+| `gm_server` | 7 |
+| `manager_server` | 9 |
+| `auction_server` | 0x0a |
+| `match_server` | 0x0b |
+| `agent_server` | 0x0d |
+| `api` | 0x17 |
+
+`gate_server`, `dungeon_server`, `pvp_server`, `guild_server` and
+`lobby_server` appear in the shipped `msg_serveroption.xml`, but
+**MessageServer's loader never reads them** (O: no string compare for them).
+Other binaries are expected to parse a different subset (I).
+
+Mapping the `db_*` connection strings to schema files in `schema/`, by their
+`Initial Catalog` (S, sample config):
+
+| Element | Database |
+|---|---|
+| `db_account`, `db_web` | `RustyHearts_Auth` (`server_info.ini` points Web at `RustyHearts_Account` instead) |
+| `db_game` | `RustyHearts` |
+| `db_game_log` | `RustyHearts_Log` |
+
+### 4.2 `option\customoption.xml` — `MagiCustomOption` (O, ctor `MS:0x140062820`, parser `MS:0x140062910`)
+
+The constructor initialises 16 boolean flags to 1 (`+0x08..+0x17`), then
+parses the file:
+- **`<log>` section:** `log_account`, `log_item`, `log_pet`, `log_loginout`,
+  `agentmanagers`, `agent`, `api`, `pvp`, `serveroption` toggle per-category
+  console logging. The parser prints `LOGVIEWER Option [%d] - … View.`.
+- **`<connect>` section:** `dbc_game`, `dbc_log`, `dbc_chat`, `dbc_emergency`,
+  `auction`, `msg`, `manager`, `match`, `gm`, `gate`, `lobby`, `dungeon`,
+  `guild`, `agentmanager`, `agent` select which peers this server connects to.
+  The parser prints `Connect Option [%d] - …`.
+- **`print`:** **?**
+
+The shipped `customoption.xml` has a different shape: flat `log_*`/`dbc_*`
+elements carrying `string` and `timeout`, which look like connection strings.
+It does **not** match the `<log>`/`<connect>` structure this parser expects.
+Either it was written for another binary, such as the DBCServer variants, or
+it isn't actually consumed (**?**). Resolve this when analysing DBCServer.
+
+### 4.3 `option\total_option.xml` (O, `MS:0x14005ceb0` agent mode; `MS:0x14005c9f0` per-server apply)
+
+This is a single combined file, used in agent mode.
+- **`MS:0x14005ceb0` (agent mode)** reads `agentmanager` (peer kind 0x15),
+  `each_option`, `dns`, `code`, `guild`, `api`, `code_start`, `setting`
+  (`multi`, `lock`), `start_server`, `revive_server` and `open_channel`. It
+  also reads the debug keys `develop`, `preload`, `preload_char`,
+  `enemy_debug`, `enemy_one`, `player_debug`, `lua_text` and `lua_send`.
+- **`MS:0x14005c9f0` (per-server apply)** reads, for DBC kinds, `db`, `game`,
+  `log`, `chat` and `emergency` (each with `string`/`timeout`) plus `count`.
+  The section chosen depends on the server kind: 0x0c→`game`, 0x11→`log`, ….
+  For kind 5 (Guild) it sets the auto flag.
+
+The shipped `total_option.xml` is empty apart from its XML declaration.
+
+### 4.4 Other files referenced (S)
+
+| File | Reference |
+|---|---|
+| `option\agent.xml` | `MS:0x14012e6d0`, parsed by `MS:0x1400602b0`: `<server …>` entries and the agent peer (kind 0x0d) |
+| `option\patch.xml` | `MS:0x14012e700`; no reader traced |
+| `option\service_control.xml` | per-country auth/billing URLs and feature flags; reader not traced in MessageServer |
+| `log_control.xml`, `content_control.xml` | in `binaries/Option/`; readers unknown |
+
+## 5. Database access path
+
+```
+ Game/Guild/Msg/... server  --ProudNet RMI (SERVERS2DBC)-->  DBCServer[_LOG]  --ADO/SQLOLEDB-->  SQL Server
+     MGDBCMsgSession pool  <--ProudNet RMI (DBC2SERVERS)--
+```
+
+- **Session pools (O, `MS:0x140004f00` startup step `DB_LOAD`):** MessageServer
+  builds four pools of `MGDBCMsgSession`, one per DBC kind: 0x0c GAME, 0x11
+  LOG, 0x12 CHAT and 0x13 EMERGENCY. Each pool has `db_cache session_count`
+  sessions (`+0x144`; `MS:0x140009330`).
+  - Each session embeds an `RNP::Servers2DbcProxy`, a subclass of the
+    PIDL-generated `SERVERS2DBC_PROTOCOL::Proxy` (`MS:0x140008e20`).
+  - The main loop pumps the pools (`DBCSessionManager->Process`) and retries
+    failed connects (`RetryConnectSession`).
+- **Request protocol (O):** `SERVERS2DBC_PROTOCOL` contains **262 RMIs, IDs
+  0x2ee1–0x2fe6**. The full name→ID table, extracted from the proxy marshal
+  functions, is in `decomp/shared/rmi_from_MessageServer.tsv`.
+  - The proxy code is generated, so every client binary is expected to contain
+    the same table (I). Verify by diffing the table extracted from each binary.
+  - Name groups include `UPDATE_*` (62), `READ_*`, `SAVE_*`, `INSERT_*`,
+    `REQ_LOG_*`, `GUILD_*`, `MAIL_*`, `PET_*`, `CHARACTER_*` and
+    `REQ_CHAT_LOG` (0x2fb2).
+- **Response protocol:** `RNP::Dbc2ServersStub` has 85 slots. Its dispatcher
+  `MS:0x140039f20` **failed to decompile** (Ghidra headless timeout, and the
+  REA call timed out too). Recover it from DBCServer's `Dbc2Servers` proxy
+  instead.
+- **SQL side:** not analysed yet. Mapping RMI → stored procedure/table needs
+  DBCServer, using `schema/*.mdf` to corroborate.
+
+## 6. Other shared protocols seen in MessageServer (O)
+
+| Protocol | IDs | Direction |
+|---|---|---|
+| `SERVERS2MSG` (stub) / `MSG2SERVERS` (proxy) | 0x3a99–0x3aa0 | servers ↔ MessageServer (`decomp/MessageServer_Release_x64/evidence.md`) |
+| `SERVER2AGENT` | 12 proxy RMIs | see the TSV |
+| `AGENT2SERVER` (stub) | ? | Agent pushes ServerInfo (§3) |
+
+## 7. Open questions (priority order)
+
+1. Decompile DBCServer:
+   - how it reads its DB connection strings;
+   - how RMIs map to SQL (stored procedure names);
+   - the `Dbc2Servers` response table.
+2. Extract the `SERVERS2DBC` table from a second client binary to confirm the
+   protocol is identical across binaries.
+3. Explain the mismatch between the `customoption.xml` shape and its parser.
+4. Decode the semantics of the `send`, `auto`, `lock` and `each_option`
+   elements, and of the `api` peer.
+5. Server kinds 0x0e–0x10 and 0x14.

docs/WORKFLOW.md

@@ -0,0 +1,203 @@
+# Workflow — reverse-engineering a server binary
+
+The repeatable process used for `Agent_Release_x64.exe`. Follow it for each
+binary in `decomp/INVENTORY.md`. Ground rules, layout and tooling are in
+`AGENTS.md`. What has been learned so far is in `docs/ARCHITECTURE.md`.
+
+Every finding is recorded as **observed**, **inference** or **unknown**, with an
+Evidence ID or decomp file reference. Static analysis never establishes runtime
+behaviour.
+
+---
+
+## 0. Prerequisites (once per machine)
+
+- Ghidra is at `/opt/ghidra`, and Java 21 and Node come from `mise.toml`.
+  - REA MCP is registered project-locally in `opencode.json`.
+  - Don't reinstall any of these; repair the configuration instead.
+- **Ghidra headless fails with `UnsupportedClassVersionError` (class file
+  71.0):** its compiled-script cache was built by a newer JDK.
+  - Move `~/.config/ghidra/ghidra_<ver>_DEV/osgi/compiled-bundles` aside.
+  - Ghidra rebuilds the cache on the next run.
+
+## 1. Identify the binary (cheap, before any decompilation)
+
+```sh
+sha256sum binaries/<X>.exe
+objdump -p binaries/<X>.exe | grep -E 'DLL Name|Subsystem|MajorLinker|pdb'
+```
+
+In REA:
+1. Run `open_binary` with the original path under `binaries/`.
+2. Run `binary_overview`.
+3. Run `search_strings` (regex) for config paths, log text, protocol names and
+   class names.
+
+Record the SHA-256, linker version, PDB path, imports and procedure count in
+`decomp/<X>/evidence.md`.
+
+## 2. Determine language and toolchain
+
+Check for these C++ markers:
+- RTTI descriptors (`strings -a | grep '^\.?AV'`)
+- `__CxxFrameHandler3`
+- an `MSVCP*.dll` import
+
+All 16 servers are MSVC 2010 C++ (see `docs/ARCHITECTURE.md`). Re-check each
+new binary and record the result.
+
+## 3. Discover third-party libraries and acquire their sources (**mandatory**)
+
+Any library you can match against upstream source saves you from
+decompiling it, and gives names, struct layouts and intent that the binary
+lacks. **Every time a library is found, search for its source online before
+continuing.**
+
+### 3a. Find candidates
+
+| Signal | How |
+|---|---|
+| Imported DLLs | `objdump -p` DLL names, cross-checked against `binaries/*.dll` |
+| DLL version resources | `strings -el <dll> \| grep -A1 -E '^(FileVersion\|ProductName\|CompanyName)$'` |
+| Embedded version text | `strings -a \| grep -iE '<lib>/[0-9]\|part of OpenSSL\|libcurl/'` |
+| Statically linked C++ libraries | RTTI class prefixes (`TiXml*`, `Proud::*`, `ATL::*`) |
+| Library-specific strings | assert/error messages (`Proud::CNetServerImpl::…`, TinyXML error table) |
+| PDB paths | e.g. `…\OpenLibrary\openssl-1.0.0d\…`, which names the exact version |
+
+### 3b. Pin the version
+
+Use, in this order of preference:
+1. An explicit version string (VERSIONINFO, `"OpenSSL 1.0.0d 8 Feb 2011"`).
+2. A version in the PDB path.
+3. **Differential evidence:** download several candidate releases and compare a
+   distinctive artefact with the binary. Examples are error-string tables,
+   exported symbol lists and header constants.
+   - Example: the TinyXML error table contains `Memory allocation failed.`.
+     That entry exists in 2.5.x and was removed in 2.6, which gives 2.5.1–2.5.3.
+4. If none of these work, choose the closest era-compatible release and mark it
+   **closest, not exact**.
+
+### 3c. Search for the source
+
+Search in this order:
+1. Official release archives (project site or SourceForge; check `old/` and
+   `archeology/` paths).
+2. Official git tags (`git ls-remote --tags <repo>`), or a GitHub tag archive
+   if the release tarball no longer exists.
+3. Vendor package repositories. These often contain headers only.
+4. Clean-room equivalents (for example mingw-w64 headers for Microsoft APIs)
+   or open reimplementations (for example Wine).
+5. `gh search repos <lib>` for forks or mirrors. Treat these as unverified and
+   label them that way.
+
+Verify each download against an official checksum where one exists. Always
+record its SHA-256.
+
+### 3d. Check the licence and place the files
+
+| Licence | Location | Committed? |
+|---|---|---|
+| Open source (OpenSSL, curl/MIT, zlib, BSD, public domain) | `third-party/<lib>-<version>/` | Yes |
+| Proprietary or EULA-restricted (e.g. ProudNet) | `third-party/_local/<lib>-<version>/` | **No.** It is gitignored. Ask the user before fetching. |
+| No public source (XTrap, Apex, HASP runtime) | nothing downloaded | Record acquisition routes only |
+
+Other rules:
+- Keep upstream files unmodified. If you must change one (for example curl's
+  version stamp), record exactly what changed.
+- Read the licence before committing anything. If redistribution is
+  prohibited, it goes in `_local/` or isn't fetched.
+
+### 3e. Document it
+
+Add a row or section to `third-party/README.md` with:
+- the observed identity in the binary, with where it was seen;
+- the version and how it was pinned;
+- the source URL, tag or commit, and SHA-256;
+- match quality (exact / closest / headers only);
+- the licence;
+- any local edits.
+
+Also:
+- Mention the library in `docs/ARCHITECTURE.md` if it is statically linked.
+- Add it to the list in `AGENTS.md`.
+- If no source exists, add it to the "No public source" and "Acquisition
+  routes" tables instead.
+
+### 3f. Use it
+
+- Label library functions in the decompiled output from known exports, class
+  method tables, error strings and the matched headers.
+- Exclude labelled library code from application analysis.
+
+## 4. Decompile
+
+In REA, choose seeds from distinctive strings or imports and use
+`trace_feature` to find the functions that reference them. Then run
+`procedure_pseudo_code` / `batch_decompile` on those functions.
+
+For full coverage, run Ghidra headless with the repo scripts:
+
+```sh
+analyzeHeadless /tmp/opencode/ghproj X -import binaries/<X>.exe -overwrite \
+  -scriptPath decomp/scripts -postScript ExportDecomp.java "$PWD/decomp/<X>/full" -deleteProject
+analyzeHeadless /tmp/opencode/ghproj X -import binaries/<X>.exe -overwrite \
+  -scriptPath decomp/scripts -postScript VTableMap.java "$PWD/decomp/<X>/full/vtables.tsv" -deleteProject
+```
+
+This produces:
+- `full/functions/<addr>.c` and `full/index.tsv` from `ExportDecomp.java`;
+- `full/vtables.tsv` from `VTableMap.java`, which maps class vftable slots to
+  functions.
+
+Generate `cpp/classes.hpp` from `vtables.tsv`, then check that it compiles:
+`g++ -fsyntax-only`.
+
+## 5. Record evidence
+
+In `decomp/<X>/`:
+1. Run `export_evidence_bundle` to write `evidence-bundle.json`.
+2. Extract the REA pseudocode to `<addr>.pseudo.c` and the overview to
+   `binary_overview.json`.
+3. Write `evidence.md`: the findings table (conclusion → type → evidence), the
+   proposed function names, and the open questions.
+4. Update the status and role of the binary in `decomp/INVENTORY.md`.
+
+## 6. Lift to readable C++
+
+`decomp/<X>/cpp/*.cpp` holds hand-lifted code for the analysed functions:
+- Cite the original address on each function.
+- Mark every name as inferred.
+- Record the original bugs you notice (leaks, races).
+
+## 7. Port to Linux
+
+`src/<name>/` builds the lifted code with Meson (see `src/README.md`). Keep a
+`platform.{hpp,cpp}` layer that maps each Win32 call to its nearest POSIX
+equivalent.
+
+The per-binary `README.md` must contain:
+- a substitution table: original address → Win32 API → Linux replacement →
+  behaviour difference;
+- what is stubbed or not ported, and why.
+
+Build with `ninja -C build` and smoke-test it. A passing test proves only that
+the port works, not how the original behaves.
+
+## 8. Close
+
+Run `close_binary` when the investigation is complete. Before closing, export
+the Evidence bundle, because session annotations are temporary.
+
+---
+
+## Checklist per binary
+
+- [ ] SHA-256, imports, linker and PDB recorded
+- [ ] Language confirmed
+- [ ] **Every third-party library identified, its version pinned, its source
+      searched for and downloaded where licensing allows, and it documented in
+      `third-party/README.md`**
+- [ ] Key functions decompiled; `full/` export done
+- [ ] `evidence.md`, `INVENTORY.md`, `docs/ARCHITECTURE.md` updated
+- [ ] Lift and Linux port updated, if in scope
+- [ ] Evidence bundle exported; REA session closed