fix: retry the packagist update on transient upstream failures

[markshust]

Problem

Every package in the split matrix reaches the Notify Packagist step within the same second, so a release fires ~92 update-package calls at Packagist at once and reliably trips its limits.

On 0.8.5 eight of them came back HTTP 500 at the same instant (run 30208431432). The single un-retried curl turned that into eight failed jobs — core, health, lsp, hashing, http, notification, page-cache, mail-smtp.

Because the failure landed after Split and push, the tags were already correct in every split repo. The only real damage was a stale Packagist index — including marko/core, which every other package resolves against. Re-running the failed jobs was enough to fix it, which is the definition of a retry we should not be doing by hand.

Change

Wrap the update call in five attempts with jittered exponential backoff:

• Retried: 5xx, 429, and curl's 000 transport failure.
• Not retried: 404 still falls through to the existing create-package self-heal; 401/403 fail immediately, since no retry fixes a bad token.

The jitter matters as much as the backoff — all 92 jobs fail on the same second, so a fixed delay would just line their retries up on the same second too.

Retrying means the curl can no longer be allowed to abort the step, so the final status is now matched against the 2xx class rather than tested with -ge 400. A transport failure arrives as 000, which is numerically zero and would otherwise be waved through as success, leaving the package stale behind a green job.

Verification

Tested behaviourally against a stub Packagist rather than only by reading the diff:

Scenario	Expected	Result
succeeds first try	exit 0	update:200
two 500s then 200 (the 0.8.5 case)	exit 0	500 500 200
five 500s, retries exhausted	exit 1	500 ×5
rate-limit 429	exit 0	429 200
unregistered package	exit 0	404 → create → 200
bad token	exit 1, no retry	401
503 then 401	exit 1	503 401
connection refused	exit 1	000 after 5 attempts

That harness also turned up a smaller wart, fixed here: curl leaves the response file untouched when it cannot connect, so a failed attempt printed the previous attempt's body as its own.

tests/SplitWorkflowTest.php gains six cases covering the retry loop, the retryable classes, the jittered backoff, the preserved 404 path, the non-2xx failure check, and the response-file reset. Full composer ci is green.

Note

The workflow tests assert on the YAML's structure, matching the existing convention in that file — they pin the shape of the fix but do not execute it. The behavioural table above is what actually verifies the logic, and it was run locally rather than in CI.

🤖 Generated with Claude Code