Reclaim artifact storage after each prebuilt run

[danielhanchen]

Ports the reclaim job from unslothai/llama.cpp unsloth-prebuilt.yml, which this pipeline was copied from without it, and sets retention-days on the upload steps that had none.

Why

Nothing in this repo ever deleted a build artifact. Measured 2026-08-11: 168 live artifacts, 13.49 GiB across 27 runs, whose binaries were already published as release assets. llama.cpp has cleaned up after itself for a while; this repo inherited the pipeline before that job existed.

Two independent gaps caused it:

1. No cleanup job. llama.cpp deletes a runs artifacts once they are provably on the release; this repo had zero artifact DELETE calls in any workflow.
2. retention-days on 2 of 5 upload steps. The rest inherited the 90-day default, which is why artifacts from early July were still alive and set to expire in October.

What this adds

reclaim job (needs: [resolve, assemble], if: always()), ported with its gates intact:

• Published runs delete only artifacts whose name matches a release asset. Build children upload sd-<tag>-bin-<label> and assemble publishes <name>.zip, so the match is exact and fixed-string. Anything unmatched is KEPT.
• Only runs/$GITHUB_RUN_ID/artifacts is ever listed, so the job cannot reach a concurrent builds artifacts.
• Runs that publish nothing delete their own bundles outright. always() rather than gating on published is deliberate: a publish:false dispatch and a cancelled run are exactly the cases that leaked in llama.cpp before.
• continue-on-error: true throughout, so cleanup can never fail a published release.

assemble.outputs.published - set only after gh release edit --draft=false lands, so reclaim can tell "these are now release assets" from "nothing will ever read these".

retention-days: 7 added to the 3 upload steps that lacked it, matching the two that already set it.

Note on the full-set fallback artifact

unsloth-sd-prebuilt-<tag> (the "Upload full set (artifact fallback)" step) does not match any release asset name, so the published path keeps it - it was the single largest artifact at 1.38 GiB. That is llama.cpps conservative rule: never delete something that is not provably on the release. With retention-days: 7 it now ages out in a week instead of 90 days. If it is genuinely redundant once the release is up, deleting it explicitly would be a cheap follow-up, but that is a behaviour change I did not want to make silently.

Verification

• unsloth-sd-prebuilt.yml parses as YAML; job graph is resolve, build-unix, build-linux-cuda, build-windows, assemble, reclaim.
• reclaim has permissions: actions: write, contents: read and both steps.
• All 5 upload steps now carry retention-days.
• Asset naming confirmed against the live master-813-bfbef5b-uff6b225 release: every bundle is <artifact name>.zip.

Test plan

• Dispatch with publish:false and confirm the unpublished path deletes the runs artifacts
• Let a scheduled publishing run complete and confirm the 5 bundles are deleted and the manifest/fallback kept
• Confirm the job summary table renders