docs: add CLAUDE.md with build, test and architecture guidance (#189)

Documents the native and cross build commands as CI runs them, the CMake
knobs that matter (the hard-reset of CMAKE_C_FLAGS, BUILD_SHARED_LIBS as
the way to drop -static, IPCHW_VENDORS), the two hardware-free tests, the
release tagging scheme, and the detection flow: getchipname() through the
UART-base and /proc/cpuinfo tables, the HAL as global function pointers,
chip_generation as the dispatch key, sensor probing over possible_i2c_addrs,
and the STANDALONE_LIBRARY split between ipctool and libipchw.

Author: openipc-ai

CLAUDE.md

@@ -0,0 +1,183 @@
+# CLAUDE.md
+
+This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.
+
+## What this is
+
+`ipctool` is a single static C99 binary that runs *on* an IP camera or DVR and
+reports its hardware as YAML: SoC, board, sensor, flash layout, RAM, firmware,
+clocks. It probes hardware directly (`/dev/mem`, I2C/SPI, `/proc`, MTD), so
+almost nothing useful executes on an x86 host. The same tree also builds
+`libipchw` (a small static library exposing chip/sensor identity, see
+`include/ipchw.h`) and `ipcinfo` (`example/ipcinfo.c`), a minimal consumer of it.
+
+Shipped binaries are static musl builds for arm32 (the canonical target),
+mips32 and arm64, and are UPX-packed before release: the raw static-musl binary
+crashes at startup on legacy kernels (Linux <= 3.18, i.e. XiongMai and old
+HiSilicon SDK firmware), and the UPX stub sidesteps that.
+
+## Build
+
+Native build (compiles everything, runs the unit test; hardware paths return
+nothing on a PC):
+
+```sh
+cmake -S . -B build -DCMAKE_BUILD_TYPE=Release
+cmake --build build -j
+./build/cYAML_test          # exit 0 == all cases passed
+```
+
+Cross build, exactly as CI does it (PR check and release both use these
+toolchains):
+
+```sh
+# arm32 (HiSilicon/Goke V1..V5, XM, SigmaStar, ...): OpenIPC hi3516cv100 musleabi toolchain
+# mips32 (Ingenic):  OpenIPC ingenic-t31 musl toolchain
+# arm64 (Hi3519DV500 etc.): Bootlin aarch64--musl--stable toolchain
+export PATH=/opt/<toolchain_dir>/bin:$PATH
+cmake -S . -B build-arm -DCMAKE_C_COMPILER=arm-openipc-linux-musleabi-gcc -DCMAKE_BUILD_TYPE=Release
+cmake --build build-arm
+upx build-arm/ipctool       # match the release artefact before testing on a camera
+```
+
+Toolchain URLs and directory names are in `.github/workflows/pr-build-check.yml`.
+`build*` and `build-arm*` are gitignored.
+
+CMake knobs worth knowing:
+
+- `CMAKE_C_FLAGS` is hard-reset to `-std=gnu99` at the top of `CMakeLists.txt`,
+  so `-DCMAKE_C_FLAGS=...` on the command line does **not** survive. Add flags
+  in `CMakeLists.txt` (guarded by a cache option) instead.
+- `-DBUILD_SHARED_LIBS=ON` is what drops the global `-static`. Needed for a
+  dynamic/ASAN build (use a glibc cross toolchain that ships libasan; musl
+  toolchains do not).
+- `-DIPCHW_VENDORS=all|none|"sstar;ingenic"` selects which vendor HALs go into
+  `libipchw`. HiSilicon is always in. The `ipctool` executable always carries
+  every vendor.
+- `-DONLY_LIBRARY=ON` builds just `libipchw`. `-DSKIP_VERSION=ON` skips the
+  git-derived `version.c` (generated by `cmake/version.cmake` on every build).
+  `SKIP_FUNDING` removes the sponsorship banner from `-h`.
+- Release flags are `-static -s -Os -ffunction-sections -Wl,--gc-sections -Wextra`;
+  binary size matters, these run from tmpfs on cameras with a few MB free.
+
+## Tests and CI
+
+- `./build/cYAML_test`: the only unit-test binary. Covers the JSON-to-YAML
+  printer (`src/cjson/cYAML.c`), including UTF-8 and invalid-byte escaping.
+- `tools/test_pipeline.sh`: hardware-free end-to-end check of the sensor
+  driver extraction pipeline (`trace_segment.py` -> `trace_to_driver.py` ->
+  `gcc -fsyntax-only` -> `trace_diff.py`). Needs only python3 and gcc.
+- `.github/workflows/pr-build-check.yml`: every PR must build clean on arm32,
+  mips32 **and** arm64 and pass `test_pipeline.sh`. Keep `#ifdef __arm__` /
+  `__mips__` / `__aarch64__` guards consistent when touching arch-specific code.
+- `.github/workflows/release.yml`: push to master publishes a rolling `latest`
+  prerelease; a `v*` tag publishes a real release. Assets are `ipctool`,
+  `ipctool-mips32`, `ipctool-arm64` plus `ipcinfo*`.
+- `.github/workflows/ci-tests.yml`: pytest against real lab cameras, triggered
+  by `repository_dispatch` only. Not runnable locally.
+
+Real verification of hardware code means running the binary on a camera.
+Cameras usually lack `sftp-server`, so copy with `scp -O`; `/tmp` is tmpfs.
+
+## Formatting
+
+`.clang-format` is LLVM style with 4-space indent. `contributors.md` asks for
+the hook: `./scripts/git-pre-commit-format install`. `scripts/apply-format`
+reformats only the changed hunks of a diff.
+
+## Architecture
+
+### Detection flow
+
+1. `src/main.c` hand-dispatches subcommands (`gpio`, `reginfo`, `i2c*`/`spi*`,
+   `clocks`, `trace`, ...) *before* getopt, then handles `-c/-s/-t`, then
+   `backup`/`upload`, and with no arguments builds the full YAML report.
+2. The report is a cJSON tree: `build_yaml()` calls one `detect_*()` per
+   section (`detect_chip`, `detect_board`, `detect_ethernet`, `get_mtd_info`,
+   `detect_ram`, `detect_firmare`, `detect_sensors`, `clocks_build_json`) and
+   `cYAML_Print()` renders it. Empty sections are dropped. New output goes
+   into a cJSON object via the `ADD_PARAM*` macros in `src/tools.h`, which
+   assume a local named `j_inner`.
+3. `getchipname()` in `src/chipid.c` is the "make sure detection ran" call and
+   is memoised. It installs the generic HAL (`setup_hal_fallback()`), then
+   `hw_detect_system()` reads the UART0 base from `/proc/iomem` to recognise
+   HiSilicon/Goke and XM SoCs by register base, falling back to
+   `generic_detect_cpu()`, which walks the `manufacturers[]` table against the
+   `/proc/cpuinfo` Hardware line. Each entry is `detect_fn` + `setup_hal_fn`,
+   compiled in per architecture and per `IPCHW_VENDOR_<NAME>` macro.
+
+### HAL: global function pointers, not vtables
+
+`src/hal/common.h` declares one global function pointer per hardware operation
+(`open_i2c_sensor_fd`, `i2c_read_register`, `i2c_change_addr`,
+`hal_temperature`, `hal_totalmem`, `hal_chip_properties`, `hal_firmware_props`,
+`hal_enable_sensor_clock`, ...). `setup_hal_fallback()` fills them with generic
+`/dev/i2c-N` implementations; each vendor's `setup_hal_<vendor>()` in
+`src/hal/<vendor>.c` overrides what that SoC needs and sets two globals every
+consumer depends on:
+
+- `chip_generation`: the switch key used by `reginfo.c` (pinmux tables),
+  `clocks.c` (PLL/DDR decoders per HiSilicon family), `bootrom.c`, `ptrace.c`,
+  `watchdog.c`, `hal/hisi/ispreg.c`, etc. HiSilicon values are the `HISI_V*`
+  constants in `src/hal/hisi/hal_hisi.h`; other vendors use their own small
+  integers. `getchipfamily()` maps them to family names.
+- `possible_i2c_addrs`: the per-vendor list of (sensor family, I2C addresses)
+  that sensor probing iterates.
+
+Adding an SoC therefore means: a detect function that fills `chip_name` and
+`chip_generation`, a `setup_hal_*` that installs bus access, and then the
+per-generation tables in each subcommand that should support it. Adding a new
+vendor also needs an entry in `IPCHW_OPTIONAL_VENDORS` in `CMakeLists.txt`, an
+include in `hal/common.h`, and a guarded row in `manufacturers[]`.
+
+Sensor I2C addresses in tables are the 8-bit (write) form; the default
+`i2c_change_addr` shifts right by one for the kernel, and some HALs install
+`i2c_change_plain_addr` instead. Ingenic gates the sensor clock, which is why
+`hal_enable_sensor_clock` runs before *every* probe rather than once.
+
+### Sensors
+
+`src/sensors.c` runs `detect_possible_sensors()`: for each address in
+`possible_i2c_addrs` it calls the vendor-family probe (`detect_sony_sensor`,
+`detect_smartsens_sensor`, ...) which reads ID registers through the HAL
+pointers and fills a `sensor_ctx_t`. Several parts share IDs or latch into
+each other after WDR cycles (IMX335/IMX347/IMX415, SP2305/OV2735), so probes
+carry fingerprint logic; read the surrounding comments before reordering
+them. `getsensoridentity()`/`getsensorshort()` are the memoised, mutex-guarded
+public entry points used by `libipchw` consumers.
+
+### Boards
+
+`src/boards/common.c` holds a table of `(is_<vendor>_board, gather_<vendor>_board_info)`
+pairs; the first detector that matches *and* gathers successfully wins.
+
+### `STANDALONE_LIBRARY`
+
+`libipchw` compiles `chipid.c`, `sensors.c`, `hal/common.c`, `hal/hisi/*` and
+the selected vendor HALs with `-DSTANDALONE_LIBRARY`. Anything that touches
+cJSON or prints diagnostics in those shared files must sit inside
+`#ifndef STANDALONE_LIBRARY`, or the library build breaks.
+
+### Architecture-specific code
+
+- `ipctool trace` (`src/ptrace.c`, `src/hal/hisi/ptrace.c`) is a ptrace-based
+  syscall decoder for camera I/O; the syscall table is hard-coded for 32-bit
+  ARM EABI and the command is compiled only under `__arm__`.
+- Register access goes through `mem_reg()` in `src/tools.c`, which mmaps
+  `/dev/mem` in 64 KiB windows and falls back to `PAGE_SIZE` when the kernel
+  rejects the larger window. Kernels built with strict devmem filtering can
+  refuse some ranges entirely.
+- `src/fake_symbols.c` holds empty definitions of HiSilicon SDK audio symbols,
+  added when the Hi3518EV100 SDK was linked in; nothing in the current tree
+  references them. `src/stack.c` is a stack-protector shim and is not in the
+  build.
+
+### Host-side tooling
+
+`tools/` holds the Python post-processing for `ipctool trace`
+(`trace_segment.py`, `trace_to_driver.py`, `trace_diff.py`),
+`capture_sensor.sh` (builds ipctool for ARM and captures a trace from a
+Majestic or Sofia camera over ssh/telnet), and firmware helpers
+(`upgrade_bundle.py`, `binwalk.py`, `telnet_upload.py`). The full workflow,
+per-family decoder coverage and troubleshooting live in
+`docs/sensor-driver-extraction.md`.