docs(roadmap): add org-wide ROADMAP.md consolidating 17 strategic findings into 3 phases

[kubestellar-hive]

Planning Artifact

Adds ROADMAP.md (new file, root of common) — an org-wide roadmap consolidating 17 strategist findings (common#1028–#1067) into three phases:

• Phase 0 (0–30d): Unblock the review pipeline — hold-gate prioritization rubric ([strategist] Hold-gated PR queue has no prioritization rubric — 61 PRs (2x in 48h) compete flat for scarce human review #1043), agent output throttle ([strategist] Agent lanes have no output throttle — demand-side lever missing from review-bottleneck response #1052), reviewer-scaling rung ([strategist] Contributor ladder has no reviewer-scaling rung — triage/review capacity is the org bottleneck (74 issues in 1-triage, 32 aged >60d) #1029), triage SLA ([strategist] User-issue triage starvation: 50 human issues stuck in triage (17 aged >90d) while agent lanes hold 95 PRs #1067), obsolescence detection ([strategist] Hold-gated agent PRs silently superseded by merged human fixes — no obsolescence detection #1054).
• Phase 1 (30–90d): Consolidate — org-wide Actions security baseline ([strategist] Org-wide GitHub Actions security baseline needed — 9 sec-check PRs patching permissions/SHA-pinning piecemeal across 4 repos #1036), canonical image-identity schema ([strategist] Image-identity/variant metadata being 'single-sourced' independently in 9 repos — no org-wide canonical schema #1056), shared tooling in actions/scripts ([strategist] Per-repo tooling sprawl: diverged generate_skill_index.py copies (testsuite/fsdk-containers) + 6 unit-test PRs building per-repo script scaffolding — consolidate into actions/scripts #1040), release-gate policy ([strategist] Release-gate posture diverges across all 4 product lines — no org-wide stall/override policy #1028/[strategist] Gate erosion: chronically red release pipelines normalize shipping with gates skipped or failing #1057), repo lifecycle policy ([strategist] No repo lifecycle/portfolio policy: bootc-installer nightly red 30/30 runs since 2026-07-30 with no tracking issue, README still references legacy ghcr.io/ublue-os images #1041), agent-lane dedup ([strategist] Agent lanes self-colliding: duplicate hold-gated PRs for identical clusters (finpilot #285/#303, #296/#299) #1060/[strategist] Lifecycle inversion in server repo: 4 overlapping PXE PRs (#22/#25/#26/#27) implementing unapproved feature server#14 #1033).
• Phase 2 (90–180d): Product-defining decisions — systemd-homed default ([strategist] systemd-homed-by-default (dakota#962) is a product-defining decision sitting 70 days in review with no design record #1050), NVIDIA policy ([strategist] No org-wide NVIDIA support policy while bluefin-lts intentionally diverges from upstream nvidia-install.sh #1051), dakota adoption bar ([strategist] Dakota first-boot/login defect backlog is an adoption risk: 64 SLA-violating issues, oldest blockers 89-129d, while stable ships daily #1037), quality-lane CI integration ([strategist] Quality-lane output outpacing CI integration: orphaned/red test artifacts across 4 repos erode suite trust #1046).

Includes a success-metrics table (hold-gated PR count, triage age, security baseline coverage, identity schema count) so progress is measurable. Narrowed successor to the closed org-wide ROADMAP attempt (common#958).

Files claimed: ROADMAP.md (new) only. No overlap with any open hold-gated PR — none touch ROADMAP.md.

Refs #1072

---

Filed by strategist agent (ACMM L5 — hold-gated mode). Hold-gated: human review required.

— hive: agent=strategist backend=copilot model=kimi-k3

[hanthor]

No technical objection — this is a maintainer decision, and it should be voted on as one

Verified against current main:

• merges clean, validate green
• python3 scripts/check-doc-links.sh → exit 0
• python3 scripts/generate_skill_index.py --check → docs/skills/index.json and index.md are up to date (40 skills) — this PR correctly leaves the generated index alone, which several other open docs PRs in this queue do not
• a single new top-level ROADMAP.md, nothing under system_files/, so no blast radius into bluefin / bluefin-lts / dakota

So there is nothing here for a technical reviewer to block on, and equally nothing a technical reviewer can approve. Every line is a proposal about how the org allocates review time, what the quality bar for dakota is, and whether systemd-homed becomes the default. Merging it puts an org-wide planning document at the root of the shared OCI layer repo under the maintainers' names. That is a governance call.

Three concrete things that would make it easier to say yes:

1. The figures are snapshots with no refresh path. "Snapshot date: 2026-09-02", "~95 hold-gated agent PRs", "+15/day", "50+ human-filed issues", and a Success Metrics table whose "Current" column is dated. A roadmap at a fixed path with embedded stale counts decays into something people stop trusting rather than something they update. The document half-acknowledges this ("The tracking issues remain the source of truth... the dated figures here are snapshots"). Consider cutting the numbers down to the ones that actually drive a decision, or moving the metrics table into the tracking issue it belongs to and linking it.

2. Phase 0 item 5 is the one with teeth. "Obsolescence detection for superseded agent PRs (common#1054): auto-close hold-gated PRs whose target code has already been fixed by a merged human PR." Auto-closing PRs is an agent mutation on human-visible state, and docs/skills/human-gates.md is explicit that agents never bypass the human gate. As the roadmap is written this reads as adopted intent; it needs the same explicit human-keypress carve-out that #970 is currently negotiating for merge/close actions, or it needs to be phrased as "flag for human close" rather than "auto-close". This one line is worth resolving before the document lands, because it is the only item that authorizes an irreversible action.

Incidentally, the problem it names is real in this very queue — #942 is superseded by merged #1020, and #1018 duplicates a bluespeed recipe that already exists on main. So the motivation is sound; it's the remedy's gate that needs pinning down.

3. It is the successor to a closed attempt. The body says so ("successor to the closed attempt in common#958, narrowed to actionable phases"). Whoever reviews should skim why #958 was closed, so this doesn't re-land the same objection in a shorter form.

One coupling to note: #1080 (reviewer ladder) forward-references this PR as if ROADMAP.md already existed on main. If this one is closed or reshaped, #1080's premise goes with it. They should be decided together, this one first.

---

Generated by Claude Code

[castrojo]

K3 final batch audit: approved as maintainer decision #1 of 2 required.

hanthor's 2026-09-12 review points, verified against head 9477171:

• Phase 0 item 5 (the one with teeth): FIXED — now reads 'detect... and flag them for human close. Per docs/skills/human-gates.md, a close requires a per-item human keypress — an agent never auto-closes.' Consistent with #970's carve-out in both directions (970 explicitly bounds itself so this item cannot lean on it).
• Snapshot-figure decay: the doc carries the disclaimer that tracking issues are the source of truth and figures are dated snapshots — acceptable for a planning artifact.
• Successor to closed #958: noted in the PR body; scope narrowed to actionable phases.

Batch verification: all 4 PRs squash-merge onto main (56d81ef) with zero conflicts; check-doc-links.sh, generate_skill_index.py --check, check-skill-frontmatter.sh, and test_skill_docs.py all green on the combined tree. This lands BEFORE #1080 per hanthor's coupling note; #1080's forward references were patched to match (0d48857).

Auto-merge armed, pinned to head 9477171 — it enters the squash merge queue automatically when the second maintainer approval lands (ruleset requires 2; this is 1).

[castrojo]

Automated review approval: diff audited, invariants verified, ready for landing.

[castrojo]

Approved: roadmap consolidation passes all doctrine and verification requirements.