[docs] MILESTONES: re-grade post-merge rubrics by evidence

[trilamsr]

Summary

Follow-up to merged PR #53. Self-review found the per-rubric ☑ markers were too generous — a code merge doesn't satisfy a rubric that requires measurement, and a CI gate only satisfies a rubric if the gate is actually fail-closed. This PR audits each shipped rubric against its file:line evidence and downgrades the ones that need it.

Re-grading

M4b (failure-injection harness): top-line ☑ delivered → ☑ partial. The nccl-hang CLI shim is a stub returning ErrPending (tools/failure-inject/ncclhang/ncclhang.go); the underlying capability exists in pkg/nccl/fr_parser/synthesize.go from M11. Three rubrics flipped to ⧗:

• nccl-hang byte-identical round-trip (CLI not wired)
• nccl-hang safe-opcode-only (CLI not wired)
• determinism on amd64 + arm64 (single-arch SHA gate; cross-arch equality is carry-forward)

M5b (Helm chart): top-line stays ☑ delivered. One rubric flipped to ⧗ — the ≤5-min hero-KPI median across 10 CI runs is satisfied by a single-run ≤300s gate; 10-run aggregation is the open work.

M10 (k8s events receiver): top-line stays ☑ alpha. Overhead-budget rubric flipped to ⧗ — BenchmarkEmitOne + BenchmarkConvertOne + rusage_linux_test.go exist, but an end-to-end 1k-events/min run asserting CPU + egress + RSS budgets together is the open work.

M11 (NCCL FlightRecorder): top-line stays ☑ alpha. Five rubrics flipped to ⧗:

• nccl-hang CLI reachability (blocked on M4b carry-forward)
• 5s dump-watcher emit timing assertion
• 2.31-drift fixture + parser-diff CI gate
• overhead bench promoted from advisory to fail-closed gate
• make generate-fixtures byte-identical regen gate

M3 (reproducible-build CI): all 11 rubrics keep ☑. Every gate is wired into release.yml and fail-closed; the lack of a real v0.X.Y tag doesn't invalidate the rubrics because they gate on workflow existence, not on past published releases.

Side cleanup

The "Foundation entries (M1/M2/M4/M9) predate this convention" line in the MILESTONES preamble belonged in FOLLOWUPS per the currency rule. Moved to docs/FOLLOWUPS.md § Documentation as an opportunistic backfill item.

Test plan

• bash scripts/doc-check.sh exits 0 — unverified marker count stable at 7
• CI green
• MILESTONES.md renders correctly on GitHub
• Every re-graded rubric has its evidence-gap noted inline in italics

Note

This PR was the third commit on the (now-merged) PR #53 branch and didn't make it into the squash. Re-applying as a focused follow-up against post-merge main.

🤖 Generated with Claude Code