Phase 3 PR 55 -- Mitsubishi MELSEC Modbus TCP quirks research document. 451-line doc at docs/v2/mitsubishi.md mirroring the docs/v2/dl205.md template for the MELSEC family (Q-series + QJ71MT91, L-series + LJ71MT91, iQ-R + RJ71EN71, iQ-R built-in Ethernet, iQ-F FX5U built-in, FX3U + FX3U-ENET / FX3U-ENET-P502, FX3GE built-in). Like Siemens S7, MELSEC Modbus is a patchwork of per-site-configured add-on modules rather than a fixed firmware stack, but the MELSEC-specific traps are different enough to warrant their own document. Key findings worth flagging for the PR 58+ implementation track: (1) MODULE NAMING TRAP -- QJ71MB91 is SERIAL RTU, not TCP. The Q-series TCP module is QJ71MT91. Driver docs + config UI should surface this clearly because the confusion costs operators hours when they try to connect to an RS-232 module via Ethernet. (2) NO CANONICAL MAPPING -- every MELSEC Modbus site has a unique 'Modbus Device Assignment Parameter' block of up to 16 assignments (each binding a MELSEC device range like D0..D1023 to a Modbus-address range); the driver must treat the mapping as runtime config, not device-family profile. (3) X/Y BASE DEPENDS ON FAMILY -- Q/L/iQ-R use HEX notation for X/Y (X20 = decimal 32), FX/iQ-F use OCTAL (X20 = decimal 16, same as DL260); iQ-F has a GX Works3 project toggle that can flip this. Single biggest off-by-N source in MELSEC driver code -- driver address helper must take a family selector. (4) Word order CDAB across Q/L/iQ-R/iQ-F by default (CPU-level, not module-level) -- no user-configurable swap on the server side. FX5U's SWAP instruction is for CLIENT mode only. Driver Mitsubishi profile default must be ByteOrder.WordSwap, matching DL260 but OPPOSITE of Siemens S7. (5) D-registers are BINARY by default (opposite of DL205's BCD-by-default). FNC 18 BCD / FNC 19 BIN instructions confirm binary-by-default in the ladder. Caller must explicitly opt-in to Bcd16/Bcd32 tags when the ladder stores BCD, same pattern as DL205 but the default is inverted. (6) FX5U FIRMWARE GATE -- needs firmware >= 1.060 for native Modbus TCP server; older firmware is client-only. Surface a clear capability error on connect. (7) FX3U PORT 502 SPLIT -- the standard FX3U-ENET cannot bind port 502 (lower port range restricted on the firmware); only FX3U-ENET-P502 can. FX3U-ENET-ADP has no Modbus at all and is a common operator mis-purchase -- driver should surface 'module does not support Modbus' as a distinct error, not 'connection refused'. (8) QJ71MT91 does NOT support FC22 (Mask Write) or FC23 (Read-Write Multiple). iQ-R and iQ-F do. Driver bulk-read optimization must gate on module capability. (9) MAX CONNECTIONS -- 16 simultaneous on Q/L/iQ-R, 8 on FX5U and FX3U-ENET. (10) STOP-mode writes -- configurable on Q/L/iQ-R/iQ-F (default = accept writes even in STOP), always rejected with exception 04 on FX3U-ENET. Per-model test differentiation section names the tests Mitsubishi_QJ71MT91_*, Mitsubishi_FX5U_*, Mitsubishi_FX3U_ENET_*, with a shared Mitsubishi_Common_* fixture for CDAB-word-order + binary-not-BCD + standard-exception-codes tests. 17 cited references including primary Mitsubishi manuals (SH-080446 for QJ71MT91, JY997D56101 for FX5, SH-081259 for iQ-R Ethernet, JY997D18101 for FX3U-ENET) plus Ignition / Kepware / Fernhill / HMS third-party driver release notes. Three unconfirmed rumours flagged explicitly: iQ-R RJ71EN71 early firmware rumoured ABCD word order (no primary source), QJ71MT91 firmware < 2010-05 FC15 odd-byte-count truncation (forum report only), FX3U-ENET firmware < 1.14 out-of-order TxId echoes under load (unreproducible on bench). Pure documentation PR -- no code, no tests. Per-quirk implementation lands in PRs 58+. Research conducted 2026-04-18.

docs/v2/mitsubishi.md

@@ -0,0 +1,451 @@
+# Mitsubishi Electric MELSEC — Modbus TCP quirks
+
+Mitsubishi's MELSEC family speaks Modbus TCP through a patchwork of add-on modules
+and built-in Ethernet ports, not a single unified stack. The module names are
+confusingly similar (`QJ71MB91` is *serial* RTU, `QJ71MT91` is the TCP/IP module
+[9]; `LJ71MT91` is the L-series equivalent; `RJ71EN71` is the iQ-R Ethernet module
+with a MODBUS/TCP *slave* mode bolted on [8]; `FX3U-ENET`, `FX3U-ENET-P502`,
+`FX3U-ENET-ADP`, `FX3GE` built-in, and `FX5U` built-in are all different code
+paths) — and every one of the categories below has at least one trap a textbook
+Modbus client gets wrong: hex-numbered X/Y devices colliding with decimal Modbus
+addresses, a user-defined "device assignment" parameter block that means *no two
+sites are identical*, CDAB-vs-ABCD word order driven by how the ladder built the
+32-bit value, sub-spec FC16 caps on the older QJ71MT91, and an FX3U port-502
+licensing split that makes `FX3U-ENET` and `FX3U-ENET-P502` different SKUs.
+This document catalogues each quirk, cites primary sources, and names the
+ModbusPal integration test we'd write for it (convention from
+`docs/v2/modbus-test-plan.md`: `Mitsubishi_<model>_<behavior>`).
+
+## Models and server/client capability
+
+| Model                  | Family   | Modbus TCP server | Modbus TCP client | Source |
+|------------------------|----------|-------------------|-------------------|--------|
+| `QJ71MT91`             | MELSEC-Q | Yes (slave)       | Yes (master)      | [9]    |
+| `QJ71MB91`             | MELSEC-Q | **Serial only** — RS-232/422/485 RTU, *not TCP* | — | [1][3] |
+| `LJ71MT91`             | MELSEC-L | Yes (slave)       | Yes (master)      | [10]   |
+| `RJ71EN71` / `RnENCPU` | MELSEC iQ-R | Yes (slave)    | Yes (master)      | [8]    |
+| `RJ71C24` / `RJ71C24-R2` | MELSEC iQ-R | RTU (serial)  | RTU (serial)      | [13]   |
+| iQ-R built-in Ethernet | CPU      | Yes (slave)       | Yes (master)      | [7]    |
+| iQ-F `FX5U` built-in Ethernet | CPU | Yes, firmware ≥ 1.060 [11] | Yes | [7][11][12] |
+| `FX3U-ENET`            | FX3U bolt-on | Yes (slave), but **not on port 502** [5] | Yes | [4][5] |
+| `FX3U-ENET-P502`       | FX3U bolt-on | Yes (slave), port 502 enabled | Yes | [5] |
+| `FX3U-ENET-ADP`        | FX3U adapter | **No MODBUS** [5] | No MODBUS     | [5]    |
+| `FX3GE` built-in       | FX3GE CPU | No MODBUS (needs ENET module) [6] | No | [6] |
+| `FX3G` + `FX3U-ENET`   | FX3G | Yes via ENET module | Yes | [6] |
+
+- A common integration mistake is to buy `FX3U-ENET-ADP` expecting MODBUS —
+  that adapter speaks only MC protocol / SLMP. Our driver should surface a clear
+  capability error, not "connection refused", when the operator's device tag
+  says `FX3U-ENET-ADP` [5].
+- Older forum threads assert the FX5U is "client only" [12] — that was true on
+  firmware ≤ 1.040. Firmware 1.060 and later ship the parameter-driven MODBUS
+  TCP server built-in and need no function blocks [11].
+
+## Modbus device assignment (the parameter block)
+
+Unlike a DL260 where the CPU exposes a *fixed* V-memory-to-Modbus mapping, every
+MELSEC MODBUS-TCP module exposes a **Modbus Device Assignment Parameter** block
+that the engineer configures in GX Works2 / GX Configurator-MB / GX Works3.
+Each of the four Modbus tables (Coil, Input, Input Register, Holding Register)
+can be split into up to 16 independent "assignment" entries, each binding a
+contiguous Modbus address range to a MELSEC device head (`M0`, `D0`, `X0`,
+`Y0`, `B0`, `W0`, `SM0`, `SD0`, `R0`, etc.) and a point count [3][7][8][9].
+
+- **There is no canonical "MELSEC Modbus mapping"**. Two sites running the same
+  QJ71MT91 module can expose completely different Modbus layouts. Our driver
+  must treat the mapping as site-data (config-file-driven), not as a device
+  profile constant.
+- **Default values do exist** — both GX Configurator-MB (for Q/L series) and
+  GX Works3 (for iQ-R / iQ-F / FX5) ship a "dedicated pattern" default that is
+  applied when the engineer does not override the assignment. Per the FX5
+  MODBUS Communication manual (JY997D56101) and the QJ71MT91 manual, the FX5
+  dedicated default is [3][7][11]:
+
+  | Modbus table       | Modbus range (0-based) | MELSEC device | Head |
+  |--------------------|------------------------|---------------|------|
+  | Coil (FC01/05/15)  | 0 – 7679               | M             | M0   |
+  | Coil               | 8192 – 8959            | Y             | Y0   |
+  | Input (FC02)       | 0 – 7679               | M             | M0   |
+  | Input              | 8192 – 8959            | X             | X0   |
+  | Input Register (FC04) | 0 – 6143            | D             | D0   |
+  | Holding Register (FC03/06/16) | 0 – 6143    | D             | D0   |
+
+  This matches the widely circulated "FC03 @ 0 = D0" convention that shows up
+  in Ubidots / Ignition / AdvancedHMI integration guides [6][12].
+
+- **X/Y in the default mapping occupy a second, non-zero Modbus range** (8192+
+  on FX5; similar on Q/L/iQ-R). Driver users who expect "X0 = coil 0" will be
+  reading M0 instead. Document this clearly.
+- **Assignment-range collisions silently disable the slave.** The QJ71MT91
+  manual states explicitly that if any two of assignments 1-16 duplicate the
+  head Modbus device number, the slave function is inactive with no clear
+  error — the module just won't respond [9]. The driver probe will look like a
+  simple timeout; the site engineer has to open GX Configurator-MB to diagnose.
+
+Test names:
+`Mitsubishi_FX5U_default_mapping_coil_0_is_M0`,
+`Mitsubishi_FX5U_default_mapping_holding_0_is_D0`,
+`Mitsubishi_QJ71MT91_duplicate_assignment_head_disables_slave`.
+
+## X/Y addressing — hex on MELSEC, decimal on Modbus
+
+**MELSEC X (input) and Y (output) device numbers are hexadecimal on Q / L /
+iQ-R** and **octal** on FX / iQ-F (with a GX Works3 toggle) [14][15].
+
+- On a Q CPU, `X20` means decimal **32**, not 20. On an FX5U in default (octal)
+  mode, `X20` means decimal **16**. GX Works3 exposes a project-level option to
+  display FX5U X/Y in hex to match Q/L/iQ-R convention — the same physical
+  input is then called `X10` [14].
+- The Modbus Device Assignment Parameter block takes the *head device* as a
+  MELSEC-native number, which is interpreted in the CPU's native base
+  (hex for Q/L/iQ-R, octal for FX/iQ-F). After that, **Modbus offsets from
+  the head are plain decimal** — the module does not apply a second hex
+  conversion [3][9].
+- Example (QJ71MT91 on a Q CPU): assignment "Coil 0 = X0, 512 points" exposes
+  physical `X0` through `X1FF` (hex) as coils 0-511. A client reading coil 32
+  gets the bit `X20` (hex) — i.e. the 33rd input, not the value at "input 20"
+  that the operator wrote on the wiring diagram in decimal.
+- **Driver bug source**: if the operator's tag configuration says "read X20" and
+  the driver helpfully converts "20" to decimal 20 → coil offset 20, the
+  returned bit is actually `X14` (hex) — off by twelve. Our config layer must
+  preserve the MELSEC-native base that the site engineer sees in GX Works.
+- Timers/counters (`T`, `C`, `ST`) are always decimal in MELSEC notation.
+  Internal relays (`M`, `B`, `L`), data registers (`D`, `W`, `R`, `ZR`),
+  and special relays/registers (`SM`, `SD`) also decimal. **Only `X` and `Y`
+  (and on Q/L/iQ-R, `B` link relays and `W` link registers) use hex**, and
+  the X/Y decision is itself family-dependent [14][15].
+
+Test names:
+`Mitsubishi_Q_X_address_is_hex_X20_equals_coil_offset_32`,
+`Mitsubishi_FX5U_X_address_is_octal_X20_equals_coil_offset_16`,
+`Mitsubishi_W_link_register_is_hex_W10_equals_holding_offset_16`.
+
+## Word order for 32-bit values
+
+MELSEC stores 32-bit ladder values (`DINT`, `DWORD`, `REAL` / single-precision
+float) across **two consecutive D-registers, low word first** — i.e., `CDAB`
+when viewed as a Modbus register pair [2][6].
+
+```
+D100 (low word)  : 0xCC 0xDD   (big-endian bytes within the word)
+D101 (high word) : 0xAA 0xBB
+```
+
+A Modbus master reading D100/D101 as a `float` with default (ABCD) word order
+gets garbage. Ignition's built-in Modbus driver notes Mitsubishi as a "CDAB
+device" specifically for this reason [2].
+
+- **Q / L / iQ-R / iQ-F all agree** — this is a CPU-level convention, not a
+  module choice. Both the QJ71MT91 manual and the FX5 MODBUS Communication
+  manual describe 32-bit access by "reading the lower 16 bits from the start
+  address and the upper 16 bits from start+1" [6][11].
+- **Byte order within each register is big-endian** (Modbus standard). The
+  module does not byte-swap.
+- **Configurable?** The MODBUS modules themselves do **not** expose a word-
+  order toggle; the behavior is fixed to how the CPU laid out the value in the
+  two D-registers. If the ladder programmer used an `SWAP` instruction or a
+  union-style assignment, the word order can be whatever they made it — but
+  for values produced by the standard `D→DBL` and `FLT`/`FLT2` instructions
+  it is always CDAB [2].
+- **FX5U quirk**: the FX5 MODBUS Communication manual tells the programmer to
+  use the `SWAP` instruction *if* the remote Modbus peer requires
+  little-endian *byte* ordering (BADC) [11]. This is only relevant when the
+  FX5U is the Modbus *client*, but it confirms the FX5U's native wire layout
+  is big-endian-byte / little-endian-word (CDAB) on the server side too.
+- **Rumoured exception**: a handful of MrPLC forum threads report iQ-R
+  RJ71EN71 firmware < 1.05 returning DWORDs in `ABCD` order when accessed via
+  the built-in Ethernet port's MODBUS slave [8]. _Unconfirmed_; treat as a
+  per-site test.
+
+Test names:
+`Mitsubishi_Float32_word_order_is_CDAB`,
+`Mitsubishi_Int32_word_order_is_CDAB`,
+`Mitsubishi_FX5U_SWAP_instruction_changes_byte_order_not_word_order`.
+
+## BCD vs binary encoding
+
+**MELSEC stores integer values in D-registers as plain binary two's-complement**,
+not BCD [16]. This is the opposite of AutomationDirect DirectLOGIC, where
+V-memory defaults to BCD and the ladder must explicitly request binary.
+
+- A ladder `MOV K1234 D100` stores `0x04D2` (1234 decimal) in D100, not
+  `0x1234`. The Modbus master reads `0x04D2` and decodes it as an integer
+  directly — no BCD conversion needed [16].
+- **Timer / counter current values** (`T0` current value, `C0` count) are
+  stored in binary as word devices on Q/L/iQ-R/iQ-F. The ladder preset
+  (`K...`) is also binary [16][17].
+- **Timer / counter preset `K` operand in FX3U / earlier FX**: also binary when
+  loaded from a D-register or a `K` constant. The older A-series CPUs had BCD
+  presets on some timer types, but MELSEC-Q, L, iQ-R, iQ-F, and FX3U all use
+  binary presets by default [17].
+- The FX3U programming manual dedicates `FNC 18 BCD` and `FNC 19 BIN` to
+  explicit conversion — their existence confirms that anything in D-registers
+  that came from a `BCD` instruction output is BCD, but nothing is BCD by
+  default [17].
+- **7-segment display registers** are a common site-specific exception — many
+  ladders pack `BCD D100` into a D-register so the operator panel can drive
+  a display directly. Our driver should not assume; expose a per-tag
+  "encoding = binary | BCD" knob.
+
+Test names:
+`Mitsubishi_D_register_stores_binary_not_BCD`,
+`Mitsubishi_FX3U_timer_current_value_is_binary`.
+
+## Max registers per request
+
+From the FX5 MODBUS Communication manual Chapter 11 [11]:
+
+| FC | Name                       | FX5U (built-in) | QJ71MT91     | iQ-R (RJ71EN71 / built-in) | FX3U-ENET |
+|----|----------------------------|-----------------|--------------|-----------------------------|-----------|
+| 01 | Read Coils                 | 1-2000          | 1-2000 [9]   | 1-2000 [8]                  | 1-2000    |
+| 02 | Read Discrete Inputs       | 1-2000          | 1-2000       | 1-2000                      | 1-2000    |
+| 03 | Read Holding Registers     | **1-125**       | 1-125 [9]    | 1-125 [8]                   | 1-125     |
+| 04 | Read Input Registers       | 1-125           | 1-125        | 1-125                       | 1-125     |
+| 05 | Write Single Coil          | 1               | 1            | 1                           | 1         |
+| 06 | Write Single Register      | 1               | 1            | 1                           | 1         |
+| 0F | Write Multiple Coils       | 1-1968          | 1-1968       | 1-1968                      | 1-1968    |
+| 10 | Write Multiple Registers   | **1-123**       | 1-123        | 1-123                       | 1-123     |
+| 16 | Mask Write Register        | 1               | not supported | 1                          | not supported |
+| 17 | Read/Write Multiple Regs   | R:1-125, W:1-121 | not supported | R:1-125, W:1-121          | not supported |
+
+- **The FX5U / iQ-R native-port limits match the Modbus spec**: 125 for FC03/04,
+  123 for FC16 [11]. No sub-spec caps like DL260's 100-register ceiling.
+- **QJ71MT91 does not support FC16 (0x16, Mask Write Register) or FC17
+  (0x17, Read/Write Multiple)** — requesting them returns exception `01`
+  Illegal Function [9]. FX5U and iQ-R *do* support both.
+- **QJ71MT91 device size**: 64k points (65,536) for each of Coil / Input /
+  Input Register / Holding Register, plus up to 4086k points for Extended
+  File Register via a secondary assignment range [9].
+- **FX3U-ENET / -P502 function code list is a strict subset** of the common
+  eight (FC01/02/03/04/05/06/0F/10). FC16 and FC17 not supported [4].
+
+Test names:
+`Mitsubishi_FX5U_FC03_126_registers_returns_IllegalDataValue`,
+`Mitsubishi_FX5U_FC16_124_registers_returns_IllegalDataValue`,
+`Mitsubishi_QJ71MT91_FC16_MaskWrite_returns_IllegalFunction`,
+`Mitsubishi_QJ71MT91_FC23_ReadWrite_returns_IllegalFunction`.
+
+## Exception codes
+
+MELSEC MODBUS modules return **only the standard Modbus exception codes 01-04**;
+no proprietary exception codes are exposed on the wire [8][9][11]. Module-
+internal diagnostics (buffer-memory error codes like `7380H`) are logged but
+not returned as Modbus exceptions.
+
+| Code | Name                 | MELSEC trigger                                          |
+|------|----------------------|---------------------------------------------------------|
+| 01   | Illegal Function     | FC17 or FC16 on QJ71MT91/FX3U; FC08 (Diagnostics); FC43 |
+| 02   | Illegal Data Address | Modbus address outside any assignment range             |
+| 03   | Illegal Data Value   | Quantity out of per-FC range (see table above); odd coil-byte count |
+| 04   | Server Device Failure | See below                                              |
+
+- **04 (Server Failure) triggers on MELSEC**:
+    - CPU in STOP or PAUSE during a write to an assignment whose "Access from
+      External Device" permission is set to "Disabled in STOP" [9][11].
+      *With the default "always enabled" setting the write succeeds in STOP
+      mode* — another common trap.
+    - CPU errors (parameter error, watchdog) during any access.
+    - Assignment points to a device range that is not configured (e.g. write
+      to `D16384` when CPU D-device size is 12288).
+- **Write to a "System Area" device** (e.g., `SD` special registers that are
+  CPU-reserved read-only) returns `04`, not `02`, on QJ71MT91 and iQ-R — the
+  assignment is valid, the device exists, but the CPU rejects the write [8][9].
+- **FX3U-ENET / -P502** returns `04` on any write attempt while the CPU is in
+  STOP, regardless of permission settings — the older firmware does not
+  implement the "Access from External Device" granularity that Q/L/iQ-R/iQ-F
+  expose [4].
+- **No rumour of proprietary codes 05-0B** from MELSEC; operators sometimes
+  report "exception 0A" but those traces all came from a third-party gateway
+  sitting between the master and the MELSEC module.
+
+Test names:
+`Mitsubishi_QJ71MT91_STOP_mode_write_with_Disabled_permission_returns_ServerFailure`,
+`Mitsubishi_QJ71MT91_STOP_mode_write_with_default_permission_succeeds`,
+`Mitsubishi_SD_system_register_write_returns_ServerFailure`,
+`Mitsubishi_FX3U_STOP_mode_write_always_returns_ServerFailure`.
+
+## Connection behavior
+
+Max simultaneous Modbus TCP clients, per module [7][8][9][11]:
+
+| Model                | Max TCP connections | Port 502 | Keepalive | Source |
+|----------------------|---------------------|----------|-----------|--------|
+| `QJ71MT91`           | 16 (shared with master role) | Yes | No | [9]    |
+| `LJ71MT91`           | 16                  | Yes      | No        | [10]   |
+| iQ-R built-in / `RJ71EN71` | 16            | Yes      | Configurable (KeepAlive = ON in parameter) | [8] |
+| iQ-F `FX5U` built-in | 8                   | Yes      | Configurable | [7][11] |
+| `FX3U-ENET`          | 8 TCP, but **not port 502** | No (port < 1024 blocked) | No | [4][5] |
+| `FX3U-ENET-P502`     | 8, port 502 enabled | Yes      | No        | [5]    |
+
+- **QJ71MT91's 16 is total connections shared between slave-listen and
+  master-initiated sockets** [9]. A site that uses the same module as both
+  master to downstream VFDs and slave to upstream SCADA splits the 16 pool.
+- **FX3U-ENET port-502 gotcha**: if the engineer loads a configuration with
+  port 502 into a non-P502 ENET module, GX Works shows the download as
+  successful; on next power cycle the module enters error state and the
+  MODBUS listener never starts. This is documented on third-party FX3G
+  integration guides [6].
+- **CPU STOP → RUN transition**: does **not** drop Modbus connections on any
+  MELSEC family. Existing sockets stay open; outstanding requests during the
+  transition may see exception 04 for a few scans but then resume [8][9].
+- **CPU reset (power cycle or `SM1255` forced reset)** drops all Modbus
+  connections and the module re-listens after typically 5-10 seconds.
+- **Idle timeout**: QJ71MT91 and iQ-R have a per-connection "Alive-Check"
+  (idle timer) parameter, default 0 (disabled). If enabled, default 10 s
+  probe interval, 3 retries before close [8][9]. FX5U similar defaults.
+- **Keep-alive (TCP-level)**: only iQ-R / iQ-F expose a TCP keep-alive option
+  (parameter "KeepAlive" in the Ethernet settings); QJ71MT91 and FX3U-ENET
+  do not — so NAT/firewall idle drops require driver-side pinging.
+
+Test names:
+`Mitsubishi_QJ71MT91_17th_connection_refused`,
+`Mitsubishi_FX5U_9th_connection_refused`,
+`Mitsubishi_STOP_to_RUN_transition_preserves_socket`,
+`Mitsubishi_CPU_reset_closes_all_sockets`.
+
+## Behavioral oddities
+
+- **Transaction ID echo**: QJ71MT91 and iQ-R reliably echo the MBAP TxId on
+  every response across firmware revisions; no reports of TxId drops under
+  load [8][9]. FX3U-ENET has an older, less-tested TCP stack; at least one
+  MrPLC thread reports out-of-order TxId echoes under heavy polling on
+  firmware < 1.14 [4]. _Unconfirmed_ on current firmware.
+- **Per-connection request serialization**: all MELSEC slaves serialize
+  requests within a single TCP connection — a new request is not processed
+  until the prior response has been sent. Pipelining multiple requests on one
+  socket causes the module to queue them in buffer memory and respond in
+  order, but **the queue depth is 1** on QJ71MT91 (a second in-flight request
+  is held on the TCP receive buffer, not queued) [9]. Driver should treat
+  Mitsubishi slaves as strictly single-flight per socket.
+- **Partial-frame handling**: QJ71MT91 and iQ-R close the socket on malformed
+  MBAP length fields. FX5U resynchronises at the next valid MBAP header
+  within 100 ms but will emit an error to `SD` diagnostics [11]. Driver must
+  reconnect on half-close and replay.
+- **FX3U UDP vs TCP**: `FX3U-ENET` supports both UDP and TCP MODBUS transports;
+  UDP is lossy and reorders under load. Default is TCP. Some legacy SCADA
+  configurations pinned the module to UDP for multicast discovery — do not
+  select UDP unless the site requires it [4].
+- **Known firmware-revision variants**:
+    - QJ71MT91 ≤ firmware 10052000000 (year-month format): FC15 with coil
+      count that forces byte-count to an odd value silently truncates the
+      last coil. Fixed in later revisions [9]. _Operator-reported_.
+    - FX5U firmware < 1.060: no native MODBUS TCP server — only accessible via
+      a predefined-protocol function block hack. Firmware ≥ 1.060 ships
+      parameter-based server. Our capability probe should read `SD203`
+      (firmware version) and flag < 1.060 as unsupported for server mode [11][12].
+    - iQ-R RJ71EN71 early firmware: possible ABCD word order (rumoured,
+      unconfirmed) [8].
+- **SD (special-register) reads during assignment-parameter load**: while
+  the CPU is loading a new MODBUS device assignment parameter (~1-2 s), the
+  slave returns exception 04 Server Failure on every request. Happens after
+  a parameter write from GX Configurator-MB [9].
+- **iQ-R "Station-based block transfer" collision**: if the RJ71EN71 is also
+  running CC-Link IE Control on the same module, a MODBUS/TCP request that
+  arrives during a CCIE cyclic period is delayed to the next scan — visible
+  as jittery response time, not a failure [8].
+
+Test names:
+`Mitsubishi_QJ71MT91_single_flight_per_socket`,
+`Mitsubishi_FX5U_malformed_MBAP_resync_within_100ms`,
+`Mitsubishi_FX3U_TxId_preserved_across_burst`,
+`Mitsubishi_FX5U_firmware_below_1_060_reports_no_server_mode`.
+
+## Model-specific differences for test coverage
+
+Summary of which quirks differ per model, so test-class naming can reflect them:
+
+| Quirk                                    | QJ71MT91 | LJ71MT91 | iQ-R (RJ71EN71 / built-in) | iQ-F (FX5U) | FX3U-ENET(-P502) |
+|------------------------------------------|----------|----------|----------------------------|-------------|------------------|
+| FC16 Mask-Write supported                | No       | No       | Yes                        | Yes         | No               |
+| FC17 Read/Write Multiple supported       | No       | No       | Yes                        | Yes         | No               |
+| Max connections                          | 16       | 16       | 16                         | 8           | 8                |
+| X/Y numbering base                       | hex      | hex      | hex                        | octal (default) | octal        |
+| 32-bit word order                        | CDAB     | CDAB     | CDAB (firmware-dependent rumour of ABCD) | CDAB | CDAB    |
+| Port 502 supported                       | Yes      | Yes      | Yes                        | Yes         | P502 only        |
+| STOP-mode write permission configurable  | Yes      | Yes      | Yes                        | Yes         | No (always blocks) |
+| TCP keep-alive parameter                 | No       | No       | Yes                        | Yes         | No               |
+| Modbus device assignment — max entries   | 16       | 16       | 16                         | 16          | 8                |
+| Server via parameter (no FB)             | Yes      | Yes      | Yes                        | Yes (fw ≥ 1.060) | Yes         |
+
+- **Test file layout**: `Mitsubishi_QJ71MT91_*`, `Mitsubishi_LJ71MT91_*`,
+  `Mitsubishi_iQR_*`, `Mitsubishi_FX5U_*`, `Mitsubishi_FX3U_ENET_*`,
+  `Mitsubishi_FX3U_ENET_P502_*`. iQ-R built-in Ethernet and the RJ71EN71
+  behave identically for MODBUS/TCP slave purposes and can share a file
+  `Mitsubishi_iQR_*`.
+- **Cross-model shared tests** (word order CDAB, binary not BCD, standard
+  exception codes, 125-register FC03 cap) can live in a single
+  `Mitsubishi_Common_*` fixture.
+
+## References
+
+1. Mitsubishi Electric, *MODBUS Interface Module User's Manual — QJ71MB91*
+   (SH-080578ENG), RS-232/422/485 MODBUS RTU serial module for MELSEC-Q —
+   https://dl.mitsubishielectric.com/dl/fa/document/manual/plc/sh080578eng/sh080578engk.pdf
+2. Inductive Automation, *Ignition Modbus Driver — Mitsubishi Q / iQ-R word
+   order*, documents CDAB convention —
+   https://docs.inductiveautomation.com/docs/8.1/ignition-modules/opc-ua/drivers/modbus-v2
+   and forum discussion https://forum.inductiveautomation.com/t/modbus-tcp-device-word-byte-order/65984
+3. Mitsubishi Electric, *Programmable Controller User's Manual QJ71MB91 MODBUS
+   Interface Module*, Chapter 7 "Parameter Setting" describing the Modbus
+   Device Assignment Parameter block (assignments 1-16, head-device
+   configuration) —
+   https://www.lcautomation.com/dbdocument/29156/QJ71MB91%20Users%20manual.pdf
+4. Mitsubishi Electric, *FX3U-ENET User's Manual* (JY997D18101), Chapter on
+   MODBUS/TCP communication; function code support and connection limits —
+   https://dl.mitsubishielectric.com/dl/fa/document/manual/plc_fx/jy997d18101/jy997d18101h.pdf
+5. Venus Automation, *Mitsubishi FX3U-ENET-P502 Module — Open Port 502 for
+   Modbus TCP/IP* —
+   https://venusautomation.com.au/mitsubishi-fx3u-enet-p502-module-open-port-502-for-modbus-tcp-ip/
+   and FX3U-ENET-ADP user manual (JY997D45801), which confirms the -ADP
+   variant does not support MODBUS —
+   https://dl.mitsubishielectric.com/dl/fa/document/manual/plc_fx/jy997d45801/jy997d45801h.pdf
+6. XML Control / Ubidots integration notes, *FX3G Modbus* — port-502 trap,
+   D-register mapping default, word order reference —
+   https://sites.google.com/site/xmlcontrol/archive/fx3g-modbus
+   and https://ubidots.com/blog/mitsubishi-plc-as-modbus-tcp-server/
+7. FA Support Me, *Modbus TCP on Built-in Ethernet port in iQ-F and iQ-R* —
+   confirms 16-connection limit on iQ-R, 8 on iQ-F, parameter-driven
+   configuration via GX Works3 —
+   https://www.fasupportme.com/portal/en/kb/articles/modbus-tcp-on-build-in-ethernet-port-in-iq-f-and-iq-r-en
+8. Mitsubishi Electric, *MELSEC iQ-R Ethernet User's Manual (Application)*
+   (SH-081259ENG) and *MELSEC iQ-RJ71EN71 User's Manual* Chapter on
+   "Communications Using Modbus/TCP" —
+   https://www.allied-automation.com/wp-content/uploads/2015/02/MITSUBISHI_manual_plc_iq-r_ethernet_users.pdf
+   and https://www.manualslib.com/manual/1533351/Mitsubishi-Electric-Melsec-Iq-Rj71en71.html?page=109
+9. Mitsubishi Electric, *MODBUS/TCP Interface Module User's Manual — QJ71MT91*
+   (SH-080446ENG), exception codes page 248, device assignment parameter
+   pages 116-124, duplicate-assignment-disables-slave note —
+   https://dl.mitsubishielectric.com/dl/fa/document/manual/plc/sh080446eng/sh080446engj.pdf
+10. Mitsubishi Electric, *MELSEC-L Network Features* — LJ71MT91 documented as
+    L-series equivalent of QJ71MT91 with identical MODBUS/TCP behavior —
+    https://us.mitsubishielectric.com/fa/en/products/cnt/programmable-controllers/melsec-l-series/network/features/
+11. Mitsubishi Electric, *MELSEC iQ-F FX5 User's Manual (MODBUS Communication)*
+    (JY997D56101), Chapter 11 "Modbus/TCP Communication Specifications" —
+    function code max-quantity table, frame specification, device assignment
+    defaults —
+    https://dl.mitsubishielectric.com/dl/fa/document/manual/plcf/jy997d56101/jy997d56101h.pdf
+12. MrPLC forum, *FX5U Modbus-TCP Server (Slave)*, firmware ≥ 1.60 enables
+    native server via parameter; earlier firmware required function block —
+    https://mrplc.com/forums/topic/31883-fx5u-modbus-tcp-server-slave/
+    and Industrial Monitor Direct's "FX5U MODBUS TCP Server Workaround"
+    article (reflects older firmware behavior) —
+    https://industrialmonitordirect.com/blogs/knowledgebase/mitsubishi-fx5u-modbus-tcp-server-configuration-workaround
+13. Mitsubishi Electric, *MELSEC iQ-R MODBUS and MODBUS/TCP Reference Manual —
+    RJ71C24 / RJ71C24-R2* (BCN-P5999-1060) — RJ71C24 is serial RTU only,
+    not TCP —
+    https://dl.mitsubishielectric.com/dl/fa/document/manual/plc/bcn-p5999-1060/bcnp59991060b.pdf
+14. HMS Industrial Networks, *eWON and Mitsubishi FX5U PLC* (KB-0264-00) —
+    documents that FX5U X/Y are octal in GX Works3 but hex when viewed as a
+    Q-series PLC through eWON; the project-level hex/octal toggle —
+    https://hmsnetworks.blob.core.windows.net/www/docs/librariesprovider10/downloads-monitored/manuals/knowledge-base/kb-0264-00-en-ewon-and-mitsubishi-fx5u-plc.pdf
+15. Fernhill Software, *Mitsubishi Melsec PLC Data Address* — documents
+    hex-vs-octal device numbering split across MELSEC families —
+    https://www.fernhillsoftware.com/help/drivers/mitsubishi-melsec/data-address-format.html
+16. Inductive Automation support, *Understanding Mitsubishi PLCs* — D registers
+    store signed 16-bit binary, not BCD; DINT combines two consecutive D
+    registers —
+    https://support.inductiveautomation.com/hc/en-us/articles/16517576753165-Understanding-Mitsubishi-PLCs
+17. Mitsubishi Electric, *FXCPU Structured Programming Manual [Device &
+    Common]* (JY997D26001) — FNC 18 BCD and FNC 19 BIN explicit-conversion
+    instructions confirm binary-by-default storage —
+    https://dl.mitsubishielectric.com/dl/fa/document/manual/plc_fx/jy997d26001/jy997d26001l.pdf