Fix: replace verbose log tiers with severity levels

.claude/skills/multi-repo-qwen-setup/SKILL.md

@@ -210,7 +210,7 @@ makespan** = the "kernel run time" layer.
 For layers ②③④ — the host↔device and bind/validate spans — parse the
 `[STRACE]` host-trace markers with `strace_timing.py` (landed in
 [simpler #1177](https://github.com/hw-native-sys/simpler/pull/1177)). The
-markers are emitted at `LOG_INFO_V9` under `SIMPLER_DFX` (no new flag),
+markers are emitted at `LOG_TIMING` under `SIMPLER_DFX` (no new flag),
 so the same log captured above carries them:
 
 ```bash
@@ -364,9 +364,9 @@ strings "$BD/$SO" | grep -m1 '<your-new-log-string>'   # confirm it baked in
 
 (`.venv/lib64` is a symlink to `lib`, so the `find` covers both.) AICPU device
 logs land in `ASCEND_PROCESS_LOG_PATH/.../device-*/device-*.log` — set that var
-per run (see `running-onboard`) and note AICPU `LOG_INFO_V*` uses **inverted**
-verbosity: `v=9` is must-see (default threshold 5), `v=0` is filtered — use
-`LOG_INFO_V9` for a diagnostic you need to read back.
+per run (see `running-onboard`). Use `--log-level info` to capture `LOG_INFO`
+diagnostics, or `--log-level debug` when the more detailed `LOG_DEBUG` stream is
+also needed.
 
 ## Anti-patterns
 

conftest.py

@@ -28,15 +28,14 @@
 import time
 import typing
 
-# Make simpler's V0..V9 and NUL acceptable to pytest's `--log-level` validator.
+# Make simpler's TIMING and NUL levels acceptable to pytest's `--log-level` validator.
 # pytest does `int(getattr(logging, level.upper(), level))`, so the value must
 # exist as a module attribute on `logging` (not just registered via
 # `addLevelName`). Set both — the addLevelName side gives nice formatter output
-# (`%(levelname)s` shows `V3` instead of `Level 18`); the setattr side is what
+# (`%(levelname)s` shows `TIMING` instead of `Level 25`); the setattr side is what
 # pytest's CLI parser actually consumes.
-for _v in range(10):
-    logging.addLevelName(15 + _v, f"V{_v}")
-    setattr(logging, f"V{_v}", 15 + _v)
+logging.addLevelName(25, "TIMING")
+setattr(logging, "TIMING", 25)
 logging.addLevelName(60, "NUL")
 setattr(logging, "NUL", 60)
 # `pytest --log-level null` upcases to "NULL" before the getattr lookup, so

docs/chip-level-arch.md

@@ -114,7 +114,7 @@ runner.finalize();
 
 ```c
 // libsimpler_log.so (RTLD_GLOBAL, loaded first by the Python wrapper):
-simpler_log_init(log_level, log_info_v);              // seed HostLogger once
+simpler_log_init(log_level);                          // seed HostLogger once
 
 // host_runtime.so (RTLD_LOCAL, loaded after):
 DeviceContextHandle ctx = create_device_context();
@@ -181,13 +181,13 @@ Python test_*.py (SceneTestCase)
   └─→ ChipWorker()
        └─→ init(device_id, bins)                          # Python wrapper
             ├─→ ctypes.CDLL(libsimpler_log.so, RTLD_GLOBAL)   # once per process
-            ├─→ simpler_log_init(log_level, log_info_v) → HostLogger seeded
+            ├─→ simpler_log_init(log_level) → HostLogger seeded
             ├─→ ctypes.CDLL(libcpu_sim_context.so, RTLD_GLOBAL)  # sim only, once
             └─→ _ChipWorker.init(host_path, aicpu_path, aicore_path, device_id)  # C++
                  ├─→ dlopen(host.so, RTLD_LOCAL) → resolve C API symbols via dlsym
                  ├─→ create_device_context() → DeviceContextHandle
                  └─→ simpler_init(ctx, device_id, aicpu*, aicpu_size, aicore*, aicore_size)
-                      ├─→ (onboard) dlog_setlevel(HostLogger.level())   # before context open
+                      ├─→ (onboard) dlog_setlevel(HostLogger.cann_level())   # before context open
                       ├─→ DeviceRunner::attach_current_thread(device_id)
                       │    ├─→ rtSetDevice(device_id) on onboard
                       │    └─→ pto_cpu_sim_bind+acquire on sim

docs/dfx/device-phases.md

@@ -80,7 +80,7 @@ deployment can profile the two domains separately:
   stores) — only whether the host re-emits the readback as markers.
 * **Host** (`simpler_run` / `bind` / `runner_run` / `validate` spans): no new
   knob — they ride the compile-time `SIMPLER_HOST_STRACE` macro and the log level
-  (`LOG_INFO_V9`), so raising the log threshold drops them.
+  (`LOG_TIMING`), so raising the log threshold drops them.
 
 `RunWall` is the whole on-NPU wall (the former `RunTiming.device_wall`); it is
 emitted as the `simpler_run.runner_run.device_wall` marker, not returned.

docs/dfx/host-trace.md

@@ -16,7 +16,7 @@ children share.
 `[STRACE]` rides on the compile-time `SIMPLER_HOST_STRACE` macro (default on, in
 `src/common/task_interface/profiling_config.h` — separate from the
 `SIMPLER_DFX` gate on the device Orch/Sched markers) and is emitted at
-`LOG_INFO_V9` (the must-see INFO tier) — **no new env var or flag**. In a
+`LOG_TIMING` (the default threshold) — **no new env var or flag**. In a
 `SIMPLER_HOST_STRACE`-off build the RAII macros compile to nothing.
 
 ## Marker grammar

docs/dfx/l2-swimlane-profiling.md

@@ -678,7 +678,7 @@ in #942, a5 still uses the legacy unified shape:
 output JSON for both. On a2a3 the orch stream replaces the per-sub-step
 records folded into ORCH_SUBMIT; there is no separate shared-memory
 aggregate. The run-window envelope is emitted to device log via
-`LOG_INFO_V9 "orch_start=… orch_end=… orch_cost=…"`.
+`LOG_INFO "orch_start=… orch_end=… orch_cost=…"`.
 
 **Producer/consumer protocol on AICore (AICore-as-producer with rotation).**
 AICore writes a slim `L2SwimlaneAicoreTaskRecord` into its currently-active per-core

docs/dynamic-linking.md

@@ -97,6 +97,12 @@ because not all symbols may be referenced. Communicates with the runtime
 through a function pointer table (`PTO2RuntimeOps`), not direct symbol
 linkage.
 
+`PTO2RuntimeOps` is a binary ABI without size or version negotiation.
+Orchestration SOs and their runtime must therefore be built from the same
+simpler revision. Changing the table's field count, order, or signatures
+invalidates previously built orchestration SOs; cached or prebuilt artifacts
+must be rebuilt before they are loaded by the updated runtime.
+
 **File path collision**: all runtimes write the orch SO to
 `/var/tmp/libdevice_orch_<PID>.so`. Safe in serial execution (each task
 dlcloses before the next writes), but would conflict in parallel in-process
@@ -281,7 +287,7 @@ AICore receives only a device copy of that same `KernelArgs` payload.
 ```text
 ChipWorker.init(device_id, bins)                       # Python wrapper
   ctypes.CDLL(libsimpler_log.so, RTLD_GLOBAL)          # once per process
-  simpler_log_init(log_level, log_info_v)              seeds HostLogger before host_runtime
+  simpler_log_init(log_level)                          seeds HostLogger before host_runtime
   ctypes.CDLL(libcpu_sim_context.so, RTLD_GLOBAL)      # sim only, once per process
   _ChipWorker.init(host_path, aicpu_path, aicore_path, device_id)   # C++
     dlopen(host_runtime.so, RTLD_LOCAL)
@@ -321,15 +327,15 @@ device_worker_main(device_id)
   for each runtime_group:
     ChipWorker.init(device_id, bins)                    # Python wrapper
       ctypes.CDLL(libsimpler_log.so, RTLD_GLOBAL)       # once per process
-      simpler_log_init(log_level, log_info_v)
+      simpler_log_init(log_level)
       _ChipWorker.init(host_path, aicpu_path, aicore_path,
                        dispatcher_path, device_id)       # C++
         dlopen(host_runtime.so, RTLD_LOCAL)
         create_device_context()
         simpler_init(ctx, device_id,
                      aicpu*, aicpu_size, aicore*, aicore_size,
                      dispatcher*, dispatcher_size)
-          dlog_setlevel(HostLogger.level())               sync CANN dlog before context open
+          dlog_setlevel(HostLogger.cann_level())          sync CANN dlog before context open
           DeviceRunner::attach_current_thread(device_id)  rtSetDevice()
           DeviceRunner::set_executors(aicpu, aicore)
           DeviceRunner::set_dispatcher_binary(dispatcher)

docs/getting-started.md

@@ -241,11 +241,11 @@ Device logs written to `~/ascend/log/debug/device-<id>/`
 Both host and AICPU kernel code use the unified `LOG_*` macros from
 `common/unified_log.h`:
 
-- `LOG_INFO_V0` .. `LOG_INFO_V9`: INFO with verbosity tier (V0 most verbose,
-  V9 most must-see, V5 default)
 - `LOG_DEBUG`: Debug messages
+- `LOG_INFO`: Lifecycle and summary messages
+- `LOG_TIMING`: Stable timing markers such as `[STRACE]`
 - `LOG_WARN`: Warnings
 - `LOG_ERROR`: Error messages
 
 Threshold is configured from Python via the `simpler` logger:
-`logging.getLogger("simpler").setLevel(simpler.V3)`.
+`logging.getLogger("simpler").setLevel(simpler.TIMING)`.