📝 docs(spec): draft S-0096, the loose ends the 0.1 nights left

[Misery7100]

This drafts the engine loose ends on the 0.2 list. It is a draft for the owner to accept. All four decisions are ASSUMED, and no LOCKED row is touched.

• D-1: a halt that cites no LOCKED row escalates underspecified. Only a halt citing a LOCKED row stays locked_conflict, which amends S-0092/D-3 (ASSUMED). Such a round then gets S-0092/D-1's one retry with the whole document's scope. This is the T-0276 case from bloomery night 17. The code change is the last return in session._halt_reason.

• D-2: a red completion battery escalates at once where no review leg runs. process_lane gains a rounds flag, set from threads.enabled with record among the sources. Without rounds, a red battery escalates the completing task as blocker_finding instead of waiting for a round nobody mints. This is the S-0095 case on torve itself, and it narrows S-0093/D-3 and D-4 to repositories that run a review leg.

• D-3: --night imports fresh contracts itself. _serve runs the scan's mint before open_night. No row made the night refuse fresh contracts; it was an ordering accident. S-0079/D-6 still refuses a queue that is empty after the import.

• D-4: the T-0113 rule also pairs a module with every existing tests/test_<stem>_*.py. Measured over the corpus's 164 phases:

  Rule	Extra test files	Phases affected
  Name prefix (chosen)	56	40
  Import scan (rejected)	1,836	106

  The prefix rule catches S-0095 phase 3's test_fleet_config.py.

Q-1: should a missing prefix-named test file refuse the contract, as T-0113 does today, or only warn? Lean: refuse. Allowing a file costs a phase nothing.

Phasing: five phases, scoped to own every test that imports a module they change.

spec fmt and spec check are clean. A dry-run plan can't run on a draft, so the phasing was checked by hand: every scoped file exists, and the phases that overlap are ordered.

🤖 Generated with Claude Code

https://claude.ai/code/session_01EabbhASMvWNhC7x4K9NJMF