FAP-WEB-LLMS-FULL-READBACK-DUPLICATE-SEMANTICS-REPAIR-01: relax llms-full readback duplicate semantics

[fermatmind]

What changed

• Treat /llms-full.txt raw repeated Enneagram canonical mentions as advisory readback evidence instead of a blocking duplicate failure.
• Keep /llms.txt duplicate handling strict.
• Upload the cache ops summary artifact with if: always() so failed readbacks still produce auditable JSON/Markdown evidence.
• Update the focused workflow contract for the split duplicate semantics and always-upload behavior.

Why

The enriched /llms-full.txt feed currently has 116 unique Enneagram canonical URLs and 232 raw URL mentions, with each canonical appearing twice. That structure should not block cache-only readback as long as unique canonical count, mode, URL hygiene, and side-effect-zero checks remain strict.

Validation

• pnpm exec vitest run tests/contracts/llms-feed-cache-ops-workflow.contract.test.ts
• actionlint .github/workflows/llms-feed-cache-ops.yml
• node scripts/verify-github-action-refs.mjs --resolve
• pnpm typecheck
• git diff --check
• changed-file scope validation limited to .github/workflows/llms-feed-cache-ops.yml and tests/contracts/llms-feed-cache-ops-workflow.contract.test.ts

Deferred items

• No deployment.
• No CMS writes or llms eligibility changes.
• No Search Queue or IndexNow calls.
• fap-api ENNEAGRAM-LLMS-FULL-RELEASE-01 ledger/readback remains deferred until this workflow succeeds and uploads the cache-only ops artifact.