chore: stop the review agent from manufacturing nits (#41)

The reviewer's findings-template output pressured it to always produce a
list; once real issues were gone it padded with cosmetic nits (label
choices, inline FQNs) that the author was told to ignore. Make an empty
findings list an explicitly valid outcome, restrict Nits to ones worth a
human's time, and state that coverage is not measured by finding count.

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

Author: dfa1

.claude/agents/zstd-reviewer.md

@@ -30,4 +30,6 @@ Read `CLAUDE.md` first. Review against it strictly. Start by running `git diff`
 - No `sun.misc.Unsafe` / internal JDK APIs. No SonarQube smells.
 
 ## Output format
-Group findings by severity: **Blocker** / **Should-fix** / **Nit**. Each finding one line: `file:line — problem — fix`. If you can, verify suspicions by running `./mvnw validate` or the relevant test and quote real output. End with a verdict: APPROVE or CHANGES NEEDED. Be specific and skeptical — your job is to catch what the coder missed.
+Group findings by severity: **Blocker** / **Should-fix** / **Nit**. Each finding one line: `file:line — problem — fix`. If you can, verify suspicions by running `./mvnw validate` or the relevant test and quote real output. End with a verdict: APPROVE or CHANGES NEEDED.
+
+Be specific and skeptical about correctness — your job is to catch what the coder missed. But **do not manufacture findings to fill the template**: an empty findings list is a valid, good outcome. A clean diff should get "APPROVE — nothing material." Report **Blocker** and **Should-fix** whenever they exist; report a **Nit only when it is genuinely worth a human's time to act on** (a real readability or correctness-adjacent issue), not for cosmetic preferences, style the build already enforces, or anything you'd caveat with "acceptable as-is / leave it." If your only findings are ones you'd tell the author to ignore, omit them and approve. Coverage is not measured by finding count.