$ Sreditor · Bug List

Building Sreditor's corroborating-signals feature — the optional git-to-code correlation tools that feed extra context into a judgment — surfaced a real batch of subprocess-integration bugs: a `cross-spawn` invocation issue, an ANSI-escape-code leak into what was supposed to be clean text output, and an `ENOBUFS` buffer overflow from a subprocess producing more output than the parent expected to capture.

Each bug traced back to treating a spawned subprocess's output as a clean, bounded string by default, when it's neither — it needs explicit encoding/stripping and an explicit buffer ceiling. Judged against its own build afterward, Sreditor's own judgment engine correctly ruled this debugging work SR&ED-ineligible: real effort, but routine subprocess-integration troubleshooting, not a resolved technological uncertainty.

Testing Sreditor against Farpost's real, much larger 48-change OpenSpec history — roughly four times the size of Sreditor's own build — surfaced a second, different batch of bugs that dogfooding against Sreditor's own smaller history never touched: a token-ceiling crash in rollup, missing reasoning output from the model on some judgments, and run-to-run grouping non-determinism.

The token-ceiling crash was fixed by having code assemble the ineligible-change bucket instead of asking the model to enumerate every ineligible change itself, which had been blowing the output ceiling at Farpost's scale. The run-to-run grouping variance wasn't "fixed" so much as surfaced and documented honestly — rollup's exact grouping boundary can shift between runs while still landing on the same underlying story, worth stating plainly rather than implying full determinism.