[sreeram-11]

Fixes the ROCm qualification feed so it matches what actually ships, and stops the release notes from linking files that don't exist.

Right now the feed looks green while nothing ROCm ships. Since b12013 the Windows ROCm builds have failed every night, which drops each ROCm version, Linux included. But failed builds get no row, and version_qualified ignores missing archives, so the feed shows three passing Linux rows per run.

Changes

1. aggregate.py

   • An archive that never built now blocks its ROCm version, so version_qualified is false when the other half of the pair is missing.
   • An archive that doesn't qualify always gets overall: "fail".
   • A run with no reports at all now blocks every version instead of crashing.

2. release.yml

   • One shared ROCM_VERSIONS list for the expected archives.
   • It must stay in step with the build matrices; a comment says so.

3. release.yml

   • Release notes only link files that were actually attached (ROCm, OpenVINO, CUDA) and say which ROCm versions were left out.
   • The notes are rewritten when a later run reuses the same release.
   • This fixes the "Not Found" links on b12062.

4. release.yml

   • The release job records which ROCm archives it checked and uploaded.
   • The publish job also reads how each ROCm build job ended (needs actions: read), and builds the records itself if the release job stopped before writing them.

5. publish_qualification.py: every expected archive gets a row; a failed build gets a placeholder row that says so. New fields:

   • published: did this run ship the archive;
   • build_result: how its build job ended;
   • tier0 / tier1: each tier's result;
   • not_applicable: for checks that were never going to run, so missing now only means "should have run and didn't".

6. publish_qualification.py: runs are ordered by run_id, and run_attempt is the attempt that published the rows. A re-run of an old run no longer jumps to the top of the feed or overwrites latest/.

7. FEED.md and README.md: updated to match, with real pass and fail examples and a section on how this feed differs from lemonade-sdk/llamacpp-rocm's.

Notes

• The feed is now schema_version: 2. The first publish after merge starts a fresh index.json, reports/ and latest/; the old v1 data stays in the branch's git history.
• When to publish a release doesn't change: versions with a missing archive were already being dropped.
• This can't be fully tested on a PR, because PR runs skip the release and publish jobs. I tested it locally against real b12062 and b12003 data + the live v1 feed, and ran the release-notes script under Node.
• The Windows ROCm 10.1 build failure is a separate issue addressed in Fix Windows ROCm builds failing to find mt.exe + stop one ROCm version cancelling the others #37. This PR only makes it visible.

[sreeram-11]

@ramkrishna2910, merging this PR since I own the changes.