docs(standards): the branch model points at the inventory instead of restating it (backend#2422)

[LukasWodka]

Summary

The branch-model bullet named its own exceptions — "docs and rfcs are main-only" — one sentence after naming repo-inventory.yml as the authority and asserting "this line follows it, not the reverse."

It then drifted from that authority on both examples:

repo	this line said	inventory says
docs	main-only, not on the train	release_train: true since 2026-08-04 — develop: required, staging present
rfcs	main-only, not on the train	has a develop taking merges, measured 2026-08-22 (backend#2242 / #306)

So the enumeration is removed rather than corrected. A list of repo names in prose will drift again; a pointer to the file that is actually measured will not. That is this file's own meta-rule — "prose is only for what tooling can't judge" — applied to itself.

Why this isn't theoretical

I read this line on 2026-08-19, concluded rfcs should be main-only, created a develop there to satisfy the first bullet, and forked the repo: 9 commits on develop against 5 on main over four days, both sides editing README.md, reconciled by hand in rfcs#50.

The line that was supposed to prevent that caused it — which is the specific failure mode the org's own guidance calls out: a doc an agent reads first manufactures the bug.

Also recorded: the trap no check catches

The inventory found it and said so plainly, so it belongs here rather than nowhere:

develop_unprotected_non_train — "a develop created and left UNPROTECTED is not flagged … this correction is human-only work: no check was going to surface it."

A develop created on a non-train repo and left unprotected is invisible to the guards. Creating one to satisfy the train bullet forks the repo silently — PRs split between the new branch and the repo's existing convention, nothing promotes between them, and the heads diverge until someone reconciles by hand. Prose is the only place that can carry this, because the guard demonstrably cannot.

Blast radius — this fans out

scripts/standards-sync.py / .github/workflows/standards-sync.yml sync this file into every repo's CLAUDE.md between the org-standards markers. So this correction propagates fleet-wide on the next sync, which is the point: the stale copy in rfcs/CLAUDE.md still says "(Sole exception: the docs repo may target main.)" — a third variant, older than both the file it came from and the reality it describes.

Type

Docs / org standards.

Test plan

No code. Verified by reading the authority rather than the prose:

• repo-inventory.yml:858-880 — docs → release_train: true, develop: required, staging divergence documented
• repo-inventory.yml:1184-1202 — rfcs → release_train: false, develop exists, re-anchored to develop_unprotected_non_train on 2026-08-22, with the "no check was going to surface it" note quoted above
• The tip of develop here is fix(2242): say why rfcs is exempt, accurately, and stop one anchor describing two repos #306, "say why rfcs is exempt, accurately, and stop one anchor describing two repos" — this PR is the prose half of that same correction, which fix(2242): say why rfcs is exempt, accurately, and stop one anchor describing two repos #306 made in the inventory but not in the standards text.

Checklist

• Targets develop
• Claims verified against repo-inventory.yml, not against the prose being replaced
• Self-assigned
• One reviewer

🤖 Generated with Claude Code

---

Note

^{Cursor Bugbot is generating a summary for commit d98ff79. Configure here.}

[saadqbal]

Approving. Three lines, and it fixes the right thing in the right direction.

Both drift examples are real — I checked the one I hadn't seen myself.repo-inventory.yml gives docsrelease_train: true, develop: required, a staging: block, and a comment reading "Enrolled 2026-08-04 (backend#1420, D1 5-of-5)". So this file called docsmain-only for twenty days after it joined the train. The rfcs half I measured myself on .github#306 — it has a develop that takes merges — so the bullet was wrong about both repos it named. That's about as clean a case as you get for "restating the authority is the defect".

Pointing at the inventory instead of enumerating exceptions is the correct fix, and scoping the first bullet to "a repo on the release train" is what makes the second one land — the old text asserted a universal model and then tried to carve out of it, which is how the carve-outs went stale independently.

The third bullet is the valuable one, and I'd have wanted it even without the drift. develop_unprotected_non_train is a state no check surfaces — the inventory says so in as many words — so "if a repo appears to lack a develop, that is a fact to verify in the inventory, not a gap to fill" is exactly the instruction that stops someone helpfully creating one. Spelling out the consequence (PRs split across two heads, nothing promotes between them, divergence until a hand reconcile) is what makes it a warning rather than a rule to be argued with.

The one thing this can't do is enforce itself: a doc that says "read the inventory" still depends on people reading it. Worth knowing that .github#314's conformance audit already checks caller presence per repo, so if the branch model itself ever becomes machine-checkable, that's where it belongs — not a change for this PR.