ci: Coverage Floor — opt-in pre_measured_coverage_artifact

[topcoder1]

Summary

Adds a new optional input pre_measured_coverage_artifact to the Coverage Floor reusable. When set, the reusable downloads a coverage-percent.txt from the named artifact, parses the percentage, and skips the install + test step entirely — runs only the floor comparison + seed + sticky-comment logic.

Why

GitHub Actions hit 90% of the May 2026 budget on the topcoder1 account on 2026-05-23. The fleet audit traced the top burn to Coverage Floor + CI running the same pytest --cov invocation twice across ~20 repos. webcrawl alone: ~290 min/wk on Coverage Floor + ~287 min/wk on CI's test job, mostly the same work.

This change is the architectural fix. ~50% reduction in Coverage Floor minutes for any repo that opts in.

Design doc

Full architecture write-up + alternatives considered: proposals/2026-05-23-coverage-floor-dedup-design.md in product-center (not yet committed there — local-only for now, but the relevant excerpt:

• Caller pattern shifts from two separate workflows (ci.yml + coverage-floor.yml) to one workflow with a needs: dependency
• Chose same-workflow + needs over cross-workflow workflow_run because PR context (head SHA, PR number, base ref) is preserved natively
• Artifact format is a simple text file with the coverage percent number — language-agnostic, caller's CI computes it however

Backwards compatibility

100%. Existing ~20 fleet callers don't pass the new input → identical behavior. Self-test of this reusable (selftest/sample.py @ 99% coverage) also unaffected. Defaults preserved everywhere.

What changes in this PR

• New input pre_measured_coverage_artifact: string (default empty)
• New step Download pre-measured coverage artifact gated if: input != ''
• All language setup steps (setup-python/go/node, install uv, git cross-org config) gated to skip when input is set
• Existing Measure coverage step renamed to id measure_fresh, gated if: input == ''
• New step Read coverage from pre-measured artifact id measure_cached, gated if: input != ''
• All downstream consumers updated to ${{ steps.measure_fresh.outputs.measured || steps.measure_cached.outputs.measured }} — GHA's || returns the first truthy value, so only the running step's output is consulted

Diff: +71/-11 lines in .github/workflows/coverage-floor.yml, nothing else touched.

Follow-up PRs

• PR 3 (topcoder1/webcrawl): wires up the opt-in as the proof point. Will be filed in parallel; its first CI run will fail until this PR merges (the new input doesn't exist on @main yet).
• Fleet rollout (post-validate, ~1 week): port to remaining ~12 repos via scripts/install-coverage-floor-dedup.sh.

Side note (out of scope)

The services_postgres and services_redis declarations remain at job level (GHA doesn't support conditional services). When pre_measured is set, these still spin up but unused — ~5-10 sec overhead per run. Acceptable; could refactor later by splitting into two jobs gated on input.

Auto-merge rationale

Change to a shared reusable workflow that ~20 callers reference. Touches .github/workflows/** — high-risk per claude-author-automerge.yml policy → manual click-merge required. actionlint clean. Backwards-compatible by design.

Test plan

• actionlint passes on this branch (verified locally)
• Self-test of the reusable still passes (the pull_request trigger runs against selftest/sample.py @ 99% coverage; new if: gates evaluate to "run the existing path" because input is empty)
• After merge: PR 3 in webcrawl validates the new path end-to-end
• After 1 week of webcrawl runs: weekly Coverage Floor minute burn drops from ~290 → ~30 (≈90% reduction on that specific check; ~$9/mo on webcrawl alone)

🤖 Generated with Claude Code

[github-actions]

Coverage Floor — mode: enforce

metric	value
measured	100.0%
floor (current)	99.0%
target	100.0%
last bumped	2026-05-12

[claude]

No issues found. Logic is sound: skipped-step outputs are correctly falsy for the || coalescing, artifact path and shell working-directory are consistent, input validation guards the artifact content before use, and all language-setup gates are paired correctly with measure_fresh.