Optimize YoloCoder with Jev context selection and bounded Muse editing

[ianu82]

Coding turns spend extra model calls collecting context, repairing edits and declaring completion. This feature brings the complete Muse + Jev product implementation together: bounded context selection, a shorter path for small UI edits, exact literal replacements, explicit whole-task completion and finite recovery from supported provider errors.

• Jev selects relevant context; Muse performs edits. Uncertain or unavailable routing falls back to the normal coding conversation. The optimized build takes precedence over /preselect, avoiding duplicate selectors.
• Normal edits finish only after explicit completion, successful tool results, a complete response and the detected project check. Patch errors include paths and remaining-budget feedback; required continuations can receive one bounded diagnostic check.
• Preserves main's architecture notes, pattern-based reads, search pre-reading and rejected-patch logging. Rejects Windows rooted/drive-relative prefetch paths, preventing an unbounded parent-directory walk.
• Includes product regression tests and build/configuration instructions. scripts/build-jev-muse.sh selects Jev jev-1.13.0 and small-edit Muse muse-spark-1-3; configure the primary model as Muse too. Plain builds retain the normal completion/recovery changes without the new router.

Validation

GitHub CI passes on Windows, Linux and macOS; formatting, shellcheck and CodeQL pass. Local macOS go test -race ./..., go vet ./..., the optimized build and version smoke check also pass. Windows record tests now isolate USERPROFILE as well as HOME; the POSIX-only cancellation fixture is explicitly skipped on Windows. Existing path-rejection coverage exercises the Windows hang fix.

The implementation has been integrated with newer main; model-performance results for that combined version have not been re-measured. This PR contains product code, regression tests and build documentation. It does not add an evaluation harness, benchmark results or experiment artifacts.