[foxnne]

Plugin rebuild loop / Measure (Linux) had never completed: every run (38083453064, 38077858541, 38012636910, 37975678599, 37969247976) was cancelled at GitHub's 1 h job limit, which put a red X on unrelated PRs touching sdk/plugin_sdk.zig.

Why it ran for an hour

In the cancelled runs, the builds took seconds (6–18 s cold), but every --watch -fincremental edit printed timeout no rebuild within 300 s. Before that, the wait for the watcher's first pass also ran its full 300 s without saying so. That's about 20 minutes per backend × 6 optimize/backend combinations.

The watcher wasn't slow, it was dead. bench.py never checked it and deleted watch.log with the scratch profile. With the log kept (this PR's first run), the reason is:

compile exe fizzy-plugin-sdk-meta-Debug Debug native failure
error: thread 2652 panic: REX_GOTPCRELX

With -fincremental on x86_64 Linux, Zig 0.16 links with its own ELF linker. That linker panics on a relocation in the C objects in the plugin's build graph (stb, freetype, tree-sitter), so the build runner aborts on the first pass. That's a limitation in Zig, not in fizzy's build. On macOS (aarch64, LLVM + the system linker) the same commands work: an edit rebuilds in 1–2 s.

What changes

• scripts/plugin-loop/bench.py
  • A wait for a rebuild also watches the watcher. If it has exited, the wait ends at once.
  • A backend's watch section ends at its first failure instead of waiting out each remaining edit. An edit gets 60 s (was 300). The first pass, a cold build, keeps 300.
  • On failure it prints the last 40 lines of the watch log and adds a failed row to the step summary with the exit code and the first error: line.
• .github/workflows/plugin-loop.yml: the job's timeout-minutes is 25 (was 60) and the Measure step gets its own 20. A hang the bench can't see now ends in 20 minutes with GitHub naming the step, against a normal run of about 6.

Measured (run 38247684393, zig 0.16.0, ubuntu-latest)

	--watch -fincremental
Debug, default / llvm / self-hosted	failed: panic: REX_GOTPCRELX (first pass)
ReleaseFast, default	16.6 / 16.3 / 16.7 s per edit
ReleaseFast, llvm	17.4 / 18.8 / 18.8 s per edit
ReleaseFast, self-hosted	failed: panic: REX_GOTPCRELX (first pass)

Measure step: 5 min 20 s; job about 6 min, green. What incremental compilation buys on Linux is now measurable on a newer Zig with zig_version = master from Actions > Run workflow.

Verification

• CI run 38247684393 on this branch: Measure (Linux) completed successfully in about 6 min, with the failed rows and watch-log tails above in its output.
• Locally (macOS, zig 0.16.0): scripts/plugin-loop/bench.py --optimize Debug --backend default --watch 2 finished in 29 s with both watch edits timed (1.2 s, 2.4 s).
• Workflow YAML parses and bench.py compiles. No Zig sources changed, so the zig build gates are unaffected.

🤖 Generated with Claude Code