Pin downgrade CI to a hosted runner

[ChrisRackauckas-Claude]

Ignore this draft until it has been reviewed by @ChrisRackauckas.

Summary

• pin the downgrade workflow to the reusable workflow's documented ubuntu-24.04 runner input
• leave the downgrade resolver, dependency floors, and test coverage unchanged

Why

The current main downgrade run (29845314639) completed resolution and build, then lost communication with its self-hosted-4vcpu-8gb runner while the test step was still active. There was no test or resolver failure in the uploaded metadata.

This is runner-specific rather than a repository regression:

• hosted run 29541435021 passed on commit 7738d76, whose tree e89ed738f45a8865b1fa74fb825af22010c80b74 is byte-identical to current main at a9adb61
• all 12 recent successful downgrade runs inspected used GitHub-hosted runners; 7 of 8 inspected failed self-hosted runs reported lost runner communication
• the exact current downgraded suite passed locally, but took 73m43.9s and reached about 6.4 GiB RSS during Gray-Scott

The centralized downgrade workflow exposes os as a public workflow_call input and documents ubuntu-24.04 as the way to force GitHub-hosted execution.

Validation

• Julia 1.10.11, julia-downgrade-compat@fab1def, exact current dependency graph: Core | 1,288,226 passed / 1,288,226 total | 73m43.9s
• actionlint .github/workflows/Downgrade.yml
• Runic 1.7 over every tracked .jl file
• git diff --check

Investigation process

This was triggered by the fleet downgrade audit. I inspected the exact Actions job and annotations, reproduced the reusable downgrade action on a clean detached checkout of a9adb61, ran the repository's official grouped test harness with the same Julia and Pkg.test options, compared the failed runner with recent hosted and self-hosted history, and used the byte-identical hosted pass as the repository boundary. No tests were skipped, silenced, loosened, or removed.

[ChrisRackauckas-Claude]

Investigation scratchpad (immutable evidence):

• PR head: 660ae97783fb208168a82442916042e3b23bbf38; base at creation: a9adb61f4c2c8db763fe53988960317dbb1a527c.
• Failed main job: run 29845314639, job 88684312799, runner ID 1965414, runner self-hosted-4vcpu-8gb-dqgwf-runner-wrl6t. Resolution and build passed; test remained in progress until the runner-lost-communication annotation after 73m06s in the test step.
• Hosted boundary: run 29541435021, job 87764327053, runner ID 1001010459, runner GitHub Actions 1001010459, passed 1,288,226/1,288,226. Its commit 7738d76f0088102f95bccfdc037d6f06935a04d5 and current main a9adb61f4c2c8db763fe53988960317dbb1a527c have byte-identical tree e89ed738f45a8865b1fa74fb825af22010c80b74 (git diff --exit-code passed).
• Exact clean-main local reproduction: Julia 1.10.11; julia-downgrade-compat@fab1defb76df9fd672f63c94df73ce131d32e134; mode deps; project .; no-promote Mooncake; exact Actions Pkg.test flags. Result: Core | 1,288,226 / 1,288,226 | 73m43.9s, exit 0.
• During Gray-Scott, the local worker was observed at about 6.4 GiB RSS / 16.4 GiB VSZ and remained compute-active. This is close to the self-hosted runner class's total 8 GiB memory.
• Branch validation: actionlint exit 0; Runic 1.7 over all tracked .jl files exit 0; git diff --check exit 0.
• Raw local test log: /home/crackauc/sandbox/tmp_20260715_055953_80193/audit/logs/finitevolumemethod-a9adb61-exact-downgrade-test-julia-1.10.11.log; SHA-256 e41cab95e588a8f75af283e7986f63d557eb7c4c05b343e0898b15425889cd7b.
• Conservative CI timer: 2026-07-21 15:32:33 EDT to 17:32:33 EDT, based on a 1h32m05s longest job on the reference PR. CI will not be polled before the timer expires.

The hosted pin is not considered verified until this PR's hosted downgrade job passes.

[ChrisRackauckas-Claude]

Post-timer hosted verification:

• @ChrisRackauckas marked this ready and merged it as e9fdb36edcdc7a8a528d47b19aaece5f5da2427e.
• Target downgrade run 29861895572, job 88740432418, ran on runner ID 1001022360 (GitHub Actions 1001022360) with label and image ubuntu-24.04.
• Result: Core | 1,288,226 / 1,288,226 | 48m36.4s; the complete job passed in 55m55s.
• The exact hosted job log was archived locally as audit/logs/finitevolumemethod-pr126-hosted-downgrade-success.log, SHA-256 b591bea145e90cc741ddfceee34e9e97b8aa1b1efc81c7901bdc76200a30b13f.

All three Core matrix legs also passed. The Julia 1 QA leg failed independently of this downgrade-only diff and is being handled as a separate clean-main investigation; it does not alter the hosted downgrade verification above.

[ChrisRackauckas-Claude]

Post-merge default-head verification:

• Current main head: e9fdb36edcdc7a8a528d47b19aaece5f5da2427e.
• Push run 29869241196, job 88765029692, passed on runner ID 1001022457 (GitHub Actions 1001022457) with label and image ubuntu-24.04.
• Result: Core | 1,288,226 / 1,288,226 | 45m39.9s; complete job duration 52m34s.
• Raw hosted log: audit/logs/finitevolumemethod-main-e9f-downgrade-success.log, SHA-256 88a3700f51d39823f5f2bc5e9cdf47f10a66e118782c662265a6919a3edfbeec.
• Job metadata: audit/logs/finitevolumemethod-main-e9f-downgrade-run-29869241196-jobs.json, SHA-256 2895db2755b1e797ba9c74ca722404526669a64b69552a7b935841148700216f.

This verifies the hosted downgrade pin both on the PR head and on the exact current default branch head.