Carve-outs via subtree split strip every commit signature (lol, #680) — procedure must choose squash-import or a ruled exception

[hyperpolymath]

Finding (Part F signing audit, 2026-09-30)

A carve-out that splits history out of a subdirectory strips every commit signature.

hyperpolymath/lol holds 8 unsigned commits on main whose subjects end in standards PR numbers (#611…#660) and which have no PR in lol. Each is a re-creation of a signed standards squash. For example, lol c4b056b has the same author, the same committer (GitHub) and the same second as standards#660's signed 1441c354. Its tree abbafb60 is byte-identical to 1441c354:lol.

So they came from the #680 carve-out ("evict a2ml, lol and the repo-guardian pair") via git subtree split or filter-repo. Rewriting a commit object necessarily drops its gpgsig.

This is one-off rather than a live producer, but it will recur on every future carve-out. Under the D215/D233 zero-bypass floor, the target repo would refuse the push outright.

Acceptance criteria

• The carve-out procedure (wherever carve-out: evict a2ml, lol and the repo-guardian pair (#490 #494 #492) #680's recipe lives) states the signature consequence and picks one of two paths:
  • (a) import into the new repo as one signed commit (history kept by a link or tag back to the standards SHA), or
  • (b) a named, owner-ruled one-time exception applied before the floor goes on.
• Any carve-out tool in the estate warns when the destination repo has required_signatures.
• No action on lol's existing history. Re-signing rewrites SHAs, and SIGNING-SETUP.md §3 forbids it.

Write-up: llm-coding-configs/claude-code/signing/PART-F-SIGNING-AUDIT.md §5d, item 4. Rulings: #787 (D254).

🤖 Generated with Claude Code