reconverge says exactly what went wrong, and we print the six lines that do not

[vyncint]

Where

crates/launchbound-prune/src/runner.rs:112-127. When cargo reconverge exits with anything other than 0 or 1, the failure is reported like this:

let stderr = String::from_utf8_lossy(&output.stderr);
let tail: String = stderr
    .lines()
    .rev()
    .take(6)          // <- the last six lines
    ...
AnalyzerOutcome::ToolError {
    detail: format!("cargo reconverge exited {exit_code}:\n{tail}"),
}

Why the last six lines are the wrong six

reconverge prints its diagnostic first and then its usage text. Measured on a real failure (cargo reconverge check --cc 80):

• 45 lines of stderr
• line 1 is the message: error: `80` is not a compute capability; expected e.g. `8.6`
• lines 2–45 are the usage text, ending with the exit-code legend

So .rev().take(6) reliably captures the legend, and reliably discards the diagnosis:

TOOL ERROR (hard stop): cargo reconverge exited 2:
                        with their recorded reasons

Exit codes: 0 = no findings at deny/confirmed confidence, 1 = findings,
2 = tool error. Findings accepted by the baseline never gate the exit
code, and their count is always reported.

It even begins mid-sentence — with their recorded reasons is the tail of --show-suppressed's description — which is why the block reads as corruption rather than as a message.

This is not specific to --cc. Any reconverge failure that ends in usage text lands here, and reconverge prints usage on every argument error.

Fix

Prefer the lines that say what went wrong. reconverge marks them: they start with error:.

let lines: Vec<&str> = stderr.lines().collect();
let errors: Vec<&str> = lines.iter().copied().filter(|l| l.starts_with("error:")).collect();
let detail = if errors.is_empty() { first_n(&lines, 6) } else { errors.join("\n") };

Falling back to the head rather than the tail when nothing matches, since a tool that prints usage puts the reason first. Worth taking the sibling issue on reconverge's side too — it should not need 44 lines of usage to report a bad flag value — but launchbound should read it correctly either way, because it cannot control what its analyzer prints.

A second thing this exposes

The same detail is repeated once per configuration. Eleven identical blocks for one cause. A tool error is a property of the invocation, not of the candidate — worth collapsing identical tool errors into one report with a count, which is also what makes the output readable at corpus scale (101 candidates).

Done when

A reconverge failure reports the line reconverge wrote to explain it, and identical failures are not repeated once per candidate.