The topic-docs contract-slice prune rule is normative but unenforced — 16 slices live on main and step 5's required check does not exist

[kyle-sexton]

docs/conventions/topic-docs/README.md:33 and its "Contract-slice lifecycle (prune with pointer)" section (README.md:290-309) are normative: the contract slice docs/topics/<slug>/ is "Committed on the task branch only; pruned before merge", durable outcomes graduate through the knowledge-vault and work-item-tracker seams, and a final commit prunes the slice leaving context pointers.

The repository does not do this.

The evidence

git ls-tree -d --name-only origin/main docs/topics/ returns 16 slices:

ai-adoption-ladder, autonomy-ignition, babysit-prs-migration, boris-video-absorption, commit-convention-well-known-path, dometrain-mcp, fable-field-guide-audit, fresh-eyes-checkpoint-audit, github-plugin-candidates, handoff-bg-cutover, ladder-climb-roadmap, loop-engineering-codification, plugin-audit-port, plugin-fleet-sync-skill, plugin-organization, underspecification

Step 5 of the lifecycle specifies the enforcement mechanism — "a required check that the net PR diff (git diff --name-only base...head) contains no path under the resolved <contract_dir>/**". That check does not exist. The docs/topics references in .github/workflows/ci.yml and scripts/docs-only-paths.txt serve docs-only PR detection, not slice pruning.

Raised by chatgpt-codex-connector against #1322, correctly, on the topic-doc slice that PR carries.

The decision this needs

Not "should #1322 prune its slice" — pruning one slice while 16 siblings stay would make the corpus less consistent, and the missing gate means nothing stops the next one either. Two coherent options:

1. Enforce the convention as written. Implement step 5's required check, then graduate and prune all 17 slices (16 on main plus docs(topics): absorb the context-engineering source into a designed, dispatched instruction-surface pass #1322's) through the vault and tracker seams. This is the larger job and it deletes design records from main that are currently readable there.
2. Amend the convention to match practice. If durable design records under docs/topics/ are actually wanted on main, the README's "task-branch-only / pruned before merge" classification is wrong and should say what the repo means. The redaction bar (README.md:315-320) would then become the load-bearing rule for what may land, rather than the prune step.

Deciding by precedent is not sufficient in either direction: 16 slices on main is evidence of what happens, not an argument for what should. But the divergence must be closed on one side, because right now the written contract and every merged PR disagree.

Not blocking #1322

#1322 lands its slice consistently with the 16 that preceded it, and its review thread is declined with a pointer here rather than resolved by a unilateral prune.

[kyle-sexton]

Autonomous work-loop lane (cycle 3, worker lane) surfaced this on the autonomous frontier
(status: needs-decision, no needs-human role label yet — same gap pattern found on #1314/#1339/#1361
in cycle 2 of this lane, and independently by another lane's cycle 9 pass).

This item is a genuine two-way policy fork, not a bug fix: (1) implement docs/topics/ step 5's
missing required check and prune all 17 slices to match the convention as written, or (2) amend
the convention's "task-branch-only / pruned before merge" classification to match the 16 slices
already merged to main. Both options are named in the issue body with their own tradeoffs (deleting
readable design records from main vs. changing what the redaction bar has to carry). Deciding by
precedent alone is explicitly called out in the issue body as insufficient.

Applying needs-human so this stops re-surfacing as an autonomous-eligible candidate on
list-frontier --autonomous. No further action from this lane — routing to whoever operates
/work-items:attend-queue for the enforce-vs-amend decision.

[kyle-sexton]

This was generated by AI, delegated directly by the repository operator to resolve this blocker.

Closing as a duplicate of #1417, which is the surviving ticket. Both describe the same defect —
docs/conventions/topic-docs/README.md:305-309 specifies a required prune check that was never
built — filed one day apart.

#1417 carries the accurate data. The true current count is 17, verified this session via
git ls-tree -d --name-only origin/main docs/topics/; this issue's 16 predates
context-engineering-claude-5 landing through #1322. #1417 also carries the dated landing table,
the ci-status needs:-list wiring note (.github/workflows/ci.yml:704), and the finding that the
convention is locally owned rather than an upstream materialization.

This issue's distinct contribution is preserved, not discarded. It named a fourth option the
other did not — amend the convention to match practice — and the decision recorded on #1417 is in
part exactly that. The convention has no representation for a program whose contract spans multiple
branches, which is what six of the seventeen directories are; #1417's decision amends
README.md:290-309 to define that shape rather than treating those six as violations. Your framing
drove that.

One correction to both issues' shared premise, recorded on #1417 in full: the lifecycle is not
uniformly unenforceable. PR #1323 executed it correctly, in full, on 2026-07-25 — graduated to
docs/adr/0004-…, routed its follow-up to #1324, pruned its slice in 22744fd664, and landed a
three-dot diff of exactly one ADR path. The convention works unassisted when a slice fits one
branch. That is what made the multi-branch-program reading the right one.

Resolution and follow-on work:

• topic-docs: the contract-slice prune gate is specified but never wired — 17 slices on main #1417 — the decision (option 3, reframed: gate now, with a bounded self-draining program-slice
  register plus a pure-deletion carve-out) and the Agent Brief for wiring the gate
• topic-docs: dispose of the 17 contract-slice directories on main — prune 8, graduate 3, register 6 #1424 — disposition of the 17 existing directories: prune 8, graduate 3, register 6

Nothing here needs a separate answer. Closing as not planned in the duplicate sense.