[dashboard] fix: align numeric validation

fix(dashboard): align numeric input validation

Author: Wangmerlyn

AGENTS.md

@@ -333,6 +333,10 @@ This file defines how coding agents should work in this repository.
   refresh/start requests should keep their modest timeout.
 - Dashboard stop-result formatting must preserve non-empty backend `message`
   fields when no stop outcome lists are populated.
+- Dashboard start-form numeric validation must stay aligned with CLI/REST
+  session creation: accept plain ASCII numeric spellings, reject leading plus
+  signs and non-ASCII/underscore digit forms, and reserve `-1` for the busy
+  threshold sentinel.
 - Public `interval` values must be finite positive seconds, including fractional seconds, capped by the Python runtime wait limit; reject `NaN`, `Infinity`, and oversized values before creating or mutating session state so keep loops cannot spin, crash, or wedge.
 - Keep human VRAM parsing centralized in `src/keep_gpu/utilities/humanized_input.py`; public integer values and digit-only strings mean bytes, public VRAM byte-equivalent values must be no more than 1 PiB, and controllers may convert valid inputs to internal tensor element counts.
 - Keep omitted public VRAM defaults aligned and low-power: CLI, Python global controller, Python single-GPU controllers, service, REST, JSON-RPC, and MCP should default to `1GiB` per GPU unless a user explicitly asks for a different reservation.

docs/guides/cli.md

@@ -153,6 +153,10 @@ The dashboard provides:
 - session creation form,
 - single-session and stop-all controls.
 
+The start form follows the same numeric token rules as the CLI: use plain ASCII
+digits, omit leading plus signs, and use `-1` only for the busy-threshold
+unconditional mode.
+
 Telemetry refresh is manual by default so an idle browser tab does not keep
 probing GPU backends. Use **Refresh Now** for a one-shot update, or enable
 **Auto refresh** for 10-second polling while the tab is visible.