# PR#93: The Error Handling Category Asks for Lost or Hidden Failures, Not for Changed Error Handling

> **WARNING:** This document describes only the change applied in this PR.
> It may already be outdated once the next PR is merged.
> Historic PR documentation is not maintained along with new PRs; treat it as a snapshot.

## Problem

The category `error_handling` of `jvm-spring-backend` asks whether a change alters error handling, retries,
or what happens after a failed operation.
That holds for almost every change of error handling, also a correct one.
Jev rated the bounded retry of a commit status in werkator PR#155 at 0.98, so the PR became YELLOW;
the Codex review rated it as not critical and found nothing.

## Non-Goals

- No change of the other categories.
- A retry that is not safe to run twice stays with the category `concurrency` (PR#60).

## Solution

- The new question asks whether the change lets a failure go unnoticed:
  an exception caught and ignored or only logged at a low level, a failure reported as a success,
  an error result no longer checked, or a partial update no longer rolled back.
- From 0.85 on, the category gives YELLOW; from 0.99 on, ORANGE, as before.
- docs/werkrevisor.werkbaum: the measurement.
- werkrevisor 0.37.0.

## Measurement

On 2026-10-12 Jev (`jev-latest`) answered the old question and two new wordings
for each main-code file of 30 merged werkator PRs (#125 to #154), of hs.hsadmin.ng #281 to #316,
werkator PR#155 and seven constructed diffs.
27 of the 36 hs.hsadmin.ng PRs were merged; the other 9 were closed without a merge and are left out.
32 of the 57 merged PRs change main code; together they have 179 main-code files.
The table shows the chosen wording.

| case | old question | new question |
|---|---|---|
| defect: an exception around a save is caught and ignored | 0.99 | 0.98 |
| defect: a failed create answers 200 OK, logged at debug | 0.98 | 0.95 |
| defect: a failed credit after a debit is only logged, no rollback | 0.97 | 0.87 |
| defect: the result of publishing a final status is ignored | 0.96 | 0.95 |
| correct: a new exception handler answers 404 | 0.93 | 0.16 |
| correct: a narrower exception type with a better message | 0.97 | 0.38 |
| correct: a changed log text | 0.05 | 0.07 |
| werkator PR#155: one retry of a commit status after an I/O error | 0.98 | 0.23 |
| 57 merged PRs, highest value | 0.95 | 0.69 |

From 0.85 on, the new question finds the four constructed defects and none of the merged PRs.
The other wording, which asks for a failure hidden from the caller, the user or the operator,
rated the missing rollback at 0.92 but the narrower exception type at 0.63.
The defects are constructed and plain, so a real one may score lower.

## Open Questions

None.