feat(push): send score explanations and trace-rule violations

[babaliauskas]

Summary

• each uploaded score row now carries the record's one-line explanation and, for trace_invariants rows that measured, violations ({source, target} of {rule_id, rule_type, tool, round_index, detail, sample}), filtered to the server's contract keys; errored rows send null
• rules_checked, owner and failure_categories are not repeated on the row (they already ship in evaluator_config / decision)
• re-vendored bundle_manifest.schema.json from the server's export (byte-identical)
• docs: DOCS.md, docs/hosted.md, docs/traces.md, llms-full.txt and the changelog say what uploads — including that trace-rule verdicts computed on imported traces upload tool names and, for arguments rules, the quoted offending value, even though the timelines stay local

Depends on

evalshift/evalshift-server PR "feat: carry score explanations and trace-rule violations" being deployed, not just merged: the server's bundle models are extra="forbid", so pushes from this branch to the current server 422 at finalize. Do not release the CLI before that deploy is live.

Until the server PR merges, tests/unit/test_bundle_shape.py::test_the_vendored_schema_matches_the_server_export fails locally when ../evalshift-server is on main (it compares against the sibling checkout); it skips in CI. Re-run it un-deselected after the server merge.

Verification

• ruff check, ruff format --check, mypy --strict, pytest -m "not integration": 2565 passed (the one environmental deselect above)
• a populated violations row validates against the vendored schema (new drift-guard test)
• pushed testApp's planner run to a local server running the server branch: the example-detail route returned the rule verdicts and the web panel rendered them

🤖 Generated with Claude Code