Gate tokens on per-device rate; require human key holds for the keyboard exemption

[cport1]

What and why

An automated client was observed on the public demo verifying invisible mode every few seconds and receiving a token each time. Two things let it through: the keyboard-only accessibility exemption accepted key events with no pointer as proof of a keyboard user, and the per-minute rate limit only produced a detection in a category that carries almost no weight.

• Per-device verification rate gate. After 10 verifications in a minute from one page instance on one device at one address, the token is withheld with reason: rate_limited. Keyed on the widget instance as well as address and fingerprint, so identical machines behind one address do not share a budget. A client that reports no instance id is not gated. Applied to Go, Node, and Python.
• Keyboard exemption checks key holds. When hold data is present, at least three holds averaging under 20ms deny the exemption, and the session is scored as pointerless. No hold data keeps the exemption, so older widgets and visitors who only tab to the checkbox are unaffected. The widget now reports an average key hold alongside its key count; durations only, never which key.
• Bench capture now holds keys with a human-like delay for the keyboard-only and screen-reader personas, so a future re-capture does not score the harness's own humans as agents.

Validation

• run-bench.js --gate: human FPR 0.00% (0/126), agent TPR 97.33%, PASS.
• Go vet and test, Node and Python suites pass, with ten new tests per server covering the exemption boundaries and the rate gate.
• Playwright accessibility, lifecycle, and PoW specs pass with the new widget.
• E2E detection suite passes against the Node server.

Compatibility

• Integrators see a new withheld reason, rate_limited, documented in the README.
• Sessions with mechanical key holds and no pointer now score as pointerless and can block.
• The device fingerprint the gate uses differs between Go and Node (canvas hash field shape), which predates this change; under mixed-runtime Redis the per-device budget is counted per runtime.

Follow-up

• Bench runs on macOS must be caffeinated; documented in bench/README.md.
• The 20ms hold threshold is set well below the panel's 41ms floor and well above automation's 1 to 8ms; a real-human capture should confirm it before it is tightened.

https://claude.ai/code/session_0198hVg9KALKecyEFDhsDxdg