[ianu82]

The merged optimization currently requires a separate binary with model linker flags, so ordinary installs skip Jev context preparation and the bounded small-edit path. Enable the complete approach automatically for MindsHub in the shared CLI/web runner, so standard builds and releases use it without special flags.

MindsHub onboarding now selects Muse 1.3 directly. Existing MindsHub connections gain Jev routing automatically; version-1 configurations using the previous mindshub_air default upgrade to Muse. Other saved model choices remain intact, and version-2 saves preserve an explicit choice of the old model. Environment connections to MindsHub also default to Muse when no model is supplied.

Recognize the actual MindsHub endpoint, including /v1 and the configured domain override, rather than trusting the provider label. Both models use the same configured endpoint and key. Other providers keep their configured coding model and normal context-reading path without automatic Jev/Muse calls. The existing completion checks, routing bounds and bounded error recovery remain intact.

Remove the special build script and replace its documentation with ordinary product behaviour. Release builds need no model flags, and automatic updates retain the defaults.

Validation:

• go test -race ./... and go vet ./... passed.
• Regression coverage starts through the public constructors without setting router/writer fields: normal and small-edit routes, context preparation, model overrides, same-key calls, project checks, and no duplicate legacy preselection.
• Configuration coverage checks endpoint recognition, environment defaults, old-default migration, explicit choices and persisted settings.
• Built the ordinary CLI using only release-style stripping/version flags. Two live MindsHub smoke checks passed: a small UI edit and a bug fix, including actual project checks. The latter loaded a version-1 configuration using the old default and successfully used Muse with Jev whole-context preparation.

This validates activation and compatibility; it is not a rerun of the full performance benchmark. No evaluation datasets, run artifacts or credentials are included.