Improve Jev evidence discovery and bounded context selection

[Stahldavid]

Problem and change

Jev context selection could discard useful local candidates, omit helpers and mistake ranking preference for sufficient evidence. This change adds balanced discovery, source-bound evidence packages, scoped usefulness gates, conditional context contributions, and preservation of the original local fallback. Missing known dependencies remain explicit partial evidence.

Includes CLI/MCP diagnostics, regression tests, offline evaluation tools, and a changeset. No learned weights are promoted automatically. Raw evaluation payloads and the local spending ledger are excluded from this PR.

Validation

• Build and workspace TypeScript checks passed.
• 391 Vitest tests passed; 7 budget/evaluator Node tests passed.
• Version consistency and whitespace checks passed.
• Final reused regression set: 27/29 positive queries contain all reference sources; negative query correctly rejected. Same 26 positive IDs improved from 21/26 with prior Jev to 25/26.
• Two intermittent misses remain (working-hours breaks and SMS non-JSON handling), both reported as partial evidence. Fresh-query model assessments did not complete cleanly. These results measure source coverage, not general answer accuracy.

The release workflow is manual; this PR does not publish npm packages.