Add outside hardware testing guide

HARDWARE_TESTING.md

@@ -0,0 +1,129 @@
+# Hardware Testing
+
+The most useful outside contribution to this project right now is **hardware
+diversity**.
+
+The existing correctness work is heavily validated on NVIDIA/Windows. Results
+from AMD, Intel, Apple, other NVIDIA architectures, different drivers, and
+different current browsers help determine which observations are general and
+which are machine-specific.
+
+## What this test is
+
+This is a local browser compute test.
+
+It does **not**:
+
+- connect to a mining pool;
+- connect to a wallet;
+- submit blocks;
+- send CAP;
+- run hidden/background mining.
+
+GPU work starts only when the tester explicitly starts a verification or
+profiling action.
+
+## Easiest test
+
+Open the deployed lab:
+
+<https://webgpu-experiment-lab.pages.dev>
+
+Use a current browser with WebGPU enabled.
+
+Run the full Core-vector correctness verification.
+
+The important result is:
+
+```text
+294 / 294
+0 mismatches
+```
+
+A failure is useful evidence too. Do not retry repeatedly until it disappears;
+record the first reproducible failure and the environment.
+
+## What to record
+
+Please record:
+
+- operating system and version;
+- browser and version;
+- GPU/adapter name;
+- GPU vendor;
+- driver version when easily available;
+- whether WebGPU was available normally or required an experimental flag;
+- full Core-vector result;
+- mismatch count;
+- pipeline/device error text, if any;
+- optional exported result JSON.
+
+Do not include account credentials, wallet data, browser profiles, or unrelated
+system logs.
+
+## Optional performance test
+
+Only after correctness passes, a tester may run the synthetic profiling or
+workgroup comparison workflows.
+
+Performance is secondary.
+
+Browser timing can vary because of:
+
+- browser scheduling;
+- CPU validation time;
+- driver behavior;
+- thermal/power state;
+- background system load.
+
+A high-variance result is valid evidence. Do not turn a noisy result into a
+claimed recommendation.
+
+## Result template
+
+```text
+Result: PASS / FAIL / INCOMPLETE
+OS:
+Browser:
+GPU:
+Vendor:
+Driver:
+WebGPU available without special flag: yes / no
+
+Core vectors:
+  matched:
+  total: 294
+  mismatches:
+
+Pipeline/device error:
+
+Optional profiling:
+  workflow:
+  samples:
+  mean throughput:
+  variability/CV:
+  recommendation produced by app: yes / no
+
+Notes:
+```
+
+## Priority hardware
+
+Especially useful environments include:
+
+1. AMD Radeon on Windows or Linux;
+2. Intel Arc / Intel integrated graphics;
+3. Apple Silicon / Safari or another WebGPU-capable browser;
+4. NVIDIA architecture other than the currently recorded Blackwell result;
+5. Linux on any supported discrete GPU.
+
+One clean correctness result from a new hardware/browser family is more useful
+than many repeated performance runs on the already-tested machine.
+
+## Pass meaning
+
+A 294/294 pass means the tested browser/GPU stack produced outputs identical to
+the committed CapStash Core vectors for this verification workload.
+
+It does **not** prove profitable mining performance, universal browser support,
+production mining safety, or compatibility with every driver version.

HARDWARE_TEST_RESULT_TEMPLATE.md

@@ -0,0 +1,33 @@
+# Hardware Test Result
+
+- Result: `PASS | FAIL | INCOMPLETE`
+- Test date (UTC):
+- Repository/deployment version or commit, if known:
+- Operating system:
+- Browser/version:
+- GPU/adapter:
+- Vendor:
+- Driver:
+- WebGPU available without special flag: `yes | no`
+
+## Correctness
+
+- Core vectors matched:
+- Core vectors total: `294`
+- Mismatches:
+- Pipeline/device error:
+
+## Optional profiling
+
+- Workflow:
+- Samples:
+- Mean throughput:
+- Median throughput:
+- Variability/CV:
+- Application recommendation: `winner | no recommendation | not run`
+
+## Notes
+
+Record reproducible browser/device behavior here. Do not include wallet
+information, account credentials, browser-profile data, or unrelated system
+logs.

README.md

@@ -158,6 +158,13 @@ loads idle with no automatic GPU workload. See [DEPLOYMENT.md](./DEPLOYMENT.md).
 - Already-submitted WebGPU dispatches cannot necessarily be interrupted, so the
   interface does not claim fake cancellation.
 
+## Outside hardware testing
+
+Results from hardware/browser families not represented in the existing evidence
+are especially useful. Start with correctness, not speed. See
+[HARDWARE_TESTING.md](./HARDWARE_TESTING.md) and
+[HARDWARE_TEST_RESULT_TEMPLATE.md](./HARDWARE_TEST_RESULT_TEMPLATE.md).
+
 ## Reproducibility and tests
 
 ```powershell