ci(build): skip test jobs a push to main does not touch, and duplicate relinker hosts

[LeandroLP23]

What

Closes #5227.

• Pushes to main: the eligibility gate reads the files of the push (compareCommitsWithBasehead, before...after) and applies the path rules that pull requests already use to relinker, register context and write watch. It falls back to running every job when it can't list the files: the first push of the branch, 300 or more files (the compare API's limit), or an API error, which is logged as a warning. The full Linux and Windows builds still run on every push.
• Relinker hosts: a new relinker-hosts output feeds the relinker matrix. When the full builds run (push and manual builds), only macOS remains, because the Linux and Windows relinker tests are subsets of the full builds' tests (see CI: a push to main runs every test job, and the Linux/Windows relinker jobs repeat tests the full builds run #5227). A manual build with the linux input off has no full Linux build, so the Linux relinker job stays in it. Automatic pull request builds, which have no full build, keep all three hosts.
• The path rules moved into one select() helper that both branches use, and are unchanged for pull requests.

Tested

• node --check on the gate script, and a local harness that runs it against a mocked context/github. It produces these outputs:

Event	full	relinker	hosts	register context	write watch
push, AGC file only	true	false	macOS	false	false
push, libc file	true	true	macOS	true	true
push, this workflow	true	true	macOS	true	true
push, 300 files / compare error / first push	true	true	macOS	true	true
manual build, relinker file	true	true	macOS	false	false
manual build, docs only	true	false	macOS	false	false
manual build, relinker file, linux: false	true	true	ubuntu+macOS	false	false
automatic PR, relinker file	false	true	ubuntu+windows+macOS	false	false

• On my fork, with this workflow on main:
  • Run 38418854722: a large push (every file differs from the fork's old main) ran every test job, with the relinker matrix reduced to macOS.
  • Run 38419688801: a docs-only push ran only eligibility, linux and windows. Relinker, register context and write watch were skipped.
  • I cancelled both full builds after the job selection, since the build and test steps are unchanged.

Checklist

• Based on current main; no other open PR implements the same functions (fix(agc): start the shader preparation threads at the first job #5196 edits a different line of this file)
• One topic per PR; follow-ups go in a new PR
• No comments in code except technical debt
• Unimplemented paths throw (NotImplemented_nid_no_patch); silent stubs are listed under Silent stubs in their library's technical debt file
• No notes, investigation .md files or images added to the repository (attach them to this PR)
• New third-party code is a submodule built from source
• Generic behaviour, not specific to one title
• Depends on: none
• Unblocks: none
• AI-assisted: yes (Claude Code)

[MarcZX10]

I found one current-main behavior to preserve before merge. This PR is based on c422b57, while current main has advanced to bf8dc61. Since this branch was created, .github/workflows/build.yml gained a workflow_dispatch boolean input linux (default true), and the Linux job honors it with inputs.linux. This PR removes that guard, so a manual run that explicitly sets linux: false would still run Linux. Please rebase/update on current main and retain the input guard while applying the push-path filtering. The current PR checks pass, but they ran before this newer manual-run behavior was added.

[LeandroLP23]

@MarcZX100 Thanks. Rebased onto main bf8dc610e. The branch never changed the Linux job's if; the guard looked dropped only because the branch predated it. The if now reads (github.event_name == 'push' || (github.event_name == 'workflow_dispatch' && inputs.linux)) && ..., the same as on main.

The new input did interact with this PR in one place, fixed in 958231f09. The relinker matrix drops the Linux host on manual builds because the full Linux build runs the same relinker tests. With linux: false that build is skipped, so the Linux relinker tests would have been lost. The gate now reads inputs.linux through the LINUX_BUILD environment variable, and a manual build with linux: false keeps the Linux relinker host (ubuntu and macOS). Pushes have no inputs and keep macOS only.

Checked the same way as before. I ran the gate script under Node against mocked contexts: the nine earlier cases are unchanged, a manual build with linux: false gives ubuntu and macOS, and a push with an empty LINUX_BUILD gives macOS. The workflow parses as YAML, and tools/check_conventions.py passes. The PR description has the new row.

[MarcZX10]

Thanks for preserving the current workflow_dispatch Linux guard and keeping the Linux relinker host for manual runs with linux: false. I verified the updated host-selection logic and the current checks now pass across Linux, Windows, and macOS relinker jobs, plus the register-context and write-watch jobs. The Linux and Windows full-build jobs are intentionally skipped by this manual run. The original review concern is resolved.

[adriadam10]

Rebased onto current main: the conflict with the multi-PR build change was in the eligibility script. select() now runs over the aggregated paths/truncated, and the linux input handling is kept. Checked the gate outputs with a mocked context/github for push (AGC file, libc file), manual with linux true/false and an automatic PR build; they match the table in the description. The Linux/Windows relinker tests are a subset of the full builds (ANYPS5_RELINKER_ONLY only adds core/relinker), so dropping those hosts when the full builds run is fine. Merging once the full Build is green.

[github-actions]

Other open pull requests touch the same code. Merging one will break the other, or both implement the same thing. If this one builds on them, list them in Depends on.

Pull request	Conflicting files (lines on main)	Same exports	Same new files
#5585	.github/workflows/build.yml L124-134, L147