fix: preserve partial column lineage diagnostics

[eitsupi]

Summary

• classify unresolved but honest column lineage as structured column_indeterminate diagnostics instead of fatal column_not_found errors
• carry column identity through cross-model renames so CLI, MCP, and impact filtering keep only diagnostics for the requested root output
• preserve the nearest known model column as the terminal when deeper row-value lineage cannot be proven
• skip redundant single-column retries after the all-column analysis already produced a scoped diagnostic

Behavior

For BigQuery row-value patterns such as ARRAY_AGG(t ...)[OFFSET(0)] AS event followed by event.field:

• 0.2.4 returned a fabricated event.field source with no diagnostic
• the 0.2.5 release candidate kept the honest nearest model terminal, but also emitted fatal column_not_found and exited 1
• this change keeps the honest nearest model terminal, emits non-fatal column_indeterminate with the root output column, and exits 0 when no fatal diagnostics remain

Diagnostics from excluded outputs and unrelated same-named upstream branches are suppressed. Missing requested columns, ambiguous columns, parse failures, and internal failures remain fatal.

Performance

The release-candidate correctness path retried the full upstream SQL once per indeterminate field. Because the SQL also grows with the number of expanded fields, downstream analysis scaled quadratically.

Linux x86_64 dist builds, --no-cache, hyperfine warmup 1 / 5 runs:

Row-value fields	Before	After
35	55.3 ± 22.4 ms	4.7 ± 0.6 ms
100	179.9 ± 42.9 ms	11.8 ± 0.9 ms
300	1.405 ± 0.169 s	30.7 ± 1.3 ms

The before and after builds produce identical stdout, stderr, and exit status for all three fixtures. The existing on-demand recovery remains enabled when the all-column analysis did not produce a matching scoped diagnostic.

Validation

• cargo test --workspace --locked
• cargo fmt --all -- --check
• cargo clippy --workspace --all-targets --all-features --locked -- -D warnings
• cargo build --profile dist --locked
• column-lineage artifact oracle and manifest/catalog validation
• synthetic N=35/100/300 semantic and performance comparison
• official 0.2.4 behavior comparison