ci: retry the compile gate on dependency-resolution failures only

[Episkey-G]

Why

ci-syntax went red on codegen PR #180, and it wasn't the code:

Unresolveable build extension: Plugin org.sonatype.central:central-publishing-maven-plugin:0.8.0
  or one of its dependencies could not be resolved
Failed to read artifact descriptor for org.sonatype.central:central-publishing-maven-plugin:jar:0.8.0

Maven simply could not fetch that plugin. Because it is declared with <extensions>true</extensions>, Maven has to resolve it during the project reading phase even for a compile-only invocation — so one network hiccup can break a gate that never needed the publishing plugin in the first place. One of the seven codegen PRs in that window hit it.

The cost isn't only a manual re-run. Auto-merge requires mergeable_state == clean, so a single flake pushes that release back to a human and fires an "auto-merge blocked" alert for what was a transient fault.

Why not -Dmaven.wagon.http.retryHandler.count

Because it does nothing here. Since Maven 3.9 the default transport is maven-resolver's HttpTransporter, not wagon — verified with mvn -X:

[DEBUG] Using transporter HttpTransporter with priority 5.0 for https://repo.maven.apache.org/maven2

The wagon property is never read. The native transport has its own retry (aether.connector.http.retryHandler.count, on by default) and #180 failed anyway, which means that attempt wasn't a retryable HTTP condition. So the retry has to live one level up, around the command.

What it does

ci-syntax now calls scripts/mvn-compile.sh, which runs mvn -B -DskipTests compile and retries only when the output matches a resolution failure (Unresolveable build extension, Failed to read artifact descriptor, Could not transfer artifact, connection resets/timeouts, …). Up to 3 attempts, backing off 10s then 30s.

A compilation error is not retried. Retrying indiscriminately would run a genuine compile error three times across 50 modules and triple the feedback time on exactly the failure you most want to hear about quickly.

Verified locally

path	result
normal compile	exit 0, 1 attempt — 51 Compiling lines over 52 modules, no Nothing to compile
real compile error	exit 2, 1 attempt, returns in 3s, logs "编译失败（非依赖解析问题），不重试"
resolution failure (mirror pointed at an unreachable host)	exit 1, 3 attempts, 44s total — the 10s + 30s backoff

The normal-compile row was checked against the stale-target/ trap: a warm build can exit 0 having compiled nothing, so the Compiling/Nothing to compile counts are quoted rather than just the exit code.