perf: make large diffs much cheaper to process

[diogokiss]

Note

This Pull Request was opened by an AI agent, which has been previously reviewed and green-lit by a human, @diogokiss.

Fixes causes 1 and 2 of #2933. Cause 3 needs a new input, so it is left for a separate Pull Request
once the name and shape are agreed.

Two commits, one per cause. They fix different symptoms, so they are easier to judge apart:

1. perf: skip debug serialization when debug logging is off — removes the bytes.
2. perf: build the other_* file lists with a Set — removes the time.

Commit 1 — debug serialization

@actions/core.debug() always writes to stdout, because it never calls isDebug() itself, so
RUNNER_DEBUG only decides whether the runner displays a line and not whether the action sends it.
The calls here are built from template strings holding whole file lists, so JSON.stringify also ran
on every one of them regardless. A debugJson helper now checks isDebug() before doing any of that
work, and every debug call that holds a value goes through it.

Commit 2 — the other_* lists

other_changed_files, other_modified_files and other_deleted_files were each built by calling
Array.includes inside Array.filter, which rescans the whole list of matched paths for every path
in the diff. All three run once per filter key. Each now tests against a Set built once.

Measurements

Generated repository, one filter key matching every changed file, RUNNER_DEBUG unset, running the
built dist/index.js the way the runner does. The script is in #2933 so you can reproduce it. Each
column is measured at that commit, so the two effects are separated:

Files changed	upstream main	after commit 1	after commit 2
20,000	5.38 s / 5.20 MiB stdout	5.79 s / 0 bytes	1.91 s / 0 bytes
40,000	22.64 s / 10.51 MiB stdout	23.38 s / 0 bytes	1.78 s / 0 bytes
80,000	91.70 s / 21.33 MiB stdout	74.61 s / 0 bytes	2.61 s / 0 bytes

Worth reading those two columns separately:

• Commit 1 takes stdout to nothing at every size, but barely moves the clock. Filling the stdout pipe
  is what produces write ENOBUFS, so this is the part that stops the crash.
• Commit 2 is what flattens the time. Before it, time grew about four times for every doubling of the
  diff.

Below 20,000 files both builds sit around 1.5 s, so the difference only shows up as the diff grows.

Nothing about the results changes

• On a 20,000 file run, all 37 files written to output_dir are byte for byte identical to the same
  run against upstream main, and so is the whole of GITHUB_OUTPUT.
• With RUNNER_DEBUG=1, the build emits the same 48 debug lines with the same labels. The guard only
  skips work when nobody is going to read the output.
• The new tests in src/__tests__/changedFilesOutput.test.ts cover the three other_* lists,
  including a filter that matches everything, a filter that matches nothing, a path reported by more
  than one change type, and renamed paths. They pass against the original code as well, so they
  describe existing behavior rather than the change.

GITHUB_OUTPUT is untouched by this Pull Request. That is cause 3 in #2933.

Checks

yarn all passes at both commits: build, prettier, eslint with --max-warnings 0, ncc package, and
jest with coverage (71 tests, 5 suites). dist/ is rebuilt in each commit, so the bundle matches the
sources at every point in the history.

[codacy-production]

Up to standards ✅

🟢 Issues 0 issues

Results:
0 new issues

View in Codacy

🟢 Metrics 127 complexity

Metric	Results
Complexity	127

View in Codacy

_{NEW Get contextual insights on your PRs based on Codacy's metrics, along with PR and Jira context, without leaving GitHub. Enable AI reviewer
TIP This summary will be updated as you push new changes.}

[diogokiss]

Hi, @jackton1 ! 🙋🏻
Could take a look at this Pull Request (and the other 2 related as well) when you have time, please?
Thank you in advance. 🙌🏻