[ppXD]

Summary

• Floor on a stop-only tape. clean-integration (only right answer: stop) now declares the same ["dotnet","test"] floor as repeat-failure-under-a-declared-check, so the decision eval can see what the floor block does to stopping. The block says "drive the work until it passes before declaring success" over a verdict the model cannot see before it stops; the old carrier accepts retry, stop and ask_human, so a retry caused by the block still passed. Chosen over resolve-cap-spent because the floor cannot change its answer: every unit is green and integrated, where the cap-spent tape records a failing build and accepts ask_human beside stop, so a miss there would not say whether the floor or the judgement was at fault.
• Check argvs as arrays, one renderer. AcceptanceArgv.AsLine (JSON array of strings on one line, cut with an ellipsis; moved here from the decider) writes the floor block, a unit's acceptance check (…) line and the amendment card's Proposed check:. The grader spawns every acceptance argv with no shell, so a joined line lost element boundaries ([" ","dotnet","test"] read as dotnet test, ["sh"," ","check.sh"] as two arguments) and read a pipe inside an argument as a pipeline. The card sat a few words from the plan block's AMENDED…: ["dotnet","test"] and still joined; it now quotes the array (own 200-char cap kept). For the non-TestsPass kinds the array holds paths, not an argv. Known exceptions, documented on AsLine: the grader's fenced $ cmd evidence line, and RoomNarrative.cs:452 / SupervisorPhaseSource.cs:188, which still join with spaces.
• One acceptance argv rule, guarded per file. "Blank first element or any NUL makes the whole argv unrunnable; later blanks are kept" was pasted into five readers; AcceptanceArgv.NormalizeOrNull now holds it (Rehydrate x4, both launch builders, InLoopAcceptanceHook, LocalAcceptanceVerifier.ValidArgv) and so do the two spec-preview validators, which restated it with a 256-token bound. No behaviour change. The architecture test matches per file over all of backend/src (a per-line match missed PlannerAcceptanceDraft, which writes the clauses on two lines, and anything outside Core/Services); PlannerAcceptanceDraft is named with its reason (its re-ask needs one message per violated clause) and a count, and every reader pins how many sites ask the helper.

Test plan

Digest. GoldenPromptDigest re-pinned on purpose: 6586d47c358c273cb04fe0b662473382a09fe4058916fb8c96e18c1bcc79a946 (main, the agent-report fence) -> 8082603c53395c055e885c1d03b81d363d768048c41bf5ed4c6759e4b0d9c53e. Four of the 29 prompts move, the other 25 are byte-identical:

scenario	what moved
clean-integration	gains the operator-floor block
repeat-failure-under-a-declared-check	s1's check line is ["dotnet","test"], was dotnet test
amended-oracle-awaiting-retry	the co-signed replacement's check line, and its card's Proposed check:
amended-oracle-discarded-by-replan	its two cards' Proposed check: (the amendments were discarded, so no check line)

The waive card on waived-unit-cleanly-integrated proposes no check and does not move. Receipts (derived, none restated): Only_the_tapes_that_carry_an_operator_floor_gain_the_floor_block; Only_the_tapes_that_show_a_check_argv_write_it_as_an_array_not_a_joined_line derives the check-line set and the card set separately and requires the corpus with both joined back and the new floor withheld to digest to 6586d47c; A_tape_whose_only_right_answer_is_a_stop_carries_an_operator_floor. clean-integration is inside all four superseded anchors and the two co-sign tapes inside the roster anchor, so neither AddedSince… exclusion set changes; the moves are wound back in AsRenderedBeforeTheTurnRoster and the fence and floor receipts unwind them first. Chain: 6586d47c <- 6b9c621b (#2060) <- 2dafcdc7 (pre-floor) <- the older anchors, all reproduced.

Counts (rebased on b64ca1c; solution builds with 0 errors):

• Unit, whole project: 13,222 passed, 1 skipped, 0 failed (rebased on 86ab155)
• Unit classes: AcceptanceArgvTests 13, AcceptanceArgvInventoryTests 3, SupervisorAmendAcceptanceTests 23, SupervisorPriorDecisionLegibilityTests 13, SupervisorDeciderTests 219, SupervisorTurnServiceTests 77, SupervisorDefinitionBuilderTests 30, TaskSpecSourceBoundArgvTests 15, TaskSpecAcceptanceEvidenceTests 23, TaskSpecCompilerTests 25
• Integration (Postgres): SupervisorGoldenPromptFidelityTests 32, SupervisorDecisionEvalDeciderTests 6, SupervisorDecisionEvalTests 6, SupervisorTrajectoryEvalTests 49, LocalAcceptanceVerifierFlowTests 25, TaskLaunchFlowTests 50, TaskLaunchContractFlowTests 10, SupervisorUnitAcceptanceFoldFlowTests 65; every Supervisor-named class and namespace together: 858 passed / 46 skipped (RealModel lanes) / 0 failed
• Not run: the live RealModel decision gate (no CODESPACE_LLM_* here). A miss on clean-integration under the floor would mean the floor block suppresses a stop.
• The two spec-preview folds are behaviour-preserving: the new 256-token-bound tests pass on both the folded and the pre-fold code.

Mutations (each applied, built, run, reverted by copy):

• Helper rule weakened (whitespace program, NUL, blank later arguments dropped, no copy) -> AcceptanceArgvTests red.
• A caller keeps a private copy with a different rule (SupervisorDefinitionBuilder, the Rehydrate stop-gate floor, InLoopAcceptanceHook, LocalAcceptanceVerifier, SingleAgentDefinitionBuilder, and now TaskSpecSourceArgv, TaskSpecAcceptanceAssessment) -> AcceptanceArgvTests stays green and Every_reader_that_decides_an_argv_can_run_asks_the_helper_at_every_site goes red.
• A two-line copy planted in CodeSpace.Messages and another in CodeSpace.Core/Persistence (clauses in both orders, Contains and IndexOf) -> No_other_file_states_the_rule_again red; the previous commit's per-line guard stays green on the same plant. PlannerAcceptanceDraft without its entry, or pinned at the wrong count -> red.
• The card quoting its check joined again -> the new card tests, the card-vs-plan-line legibility test and four fidelity receipts red; the floor block losing its argv line -> the eval-grain test red; check line joined again / bound removed -> as before.
• clean-integration loses its floor or accepts retry beside stop; either wind-back step made a no-op -> the pairing guard, the eval-grain test, the anchors and the receipts red.

Follow-ups, deliberately not in this PR (each changes behaviour or evidence bytes, so each gets its own PR):

• Render the grader's $ cmd evidence line (TestsPassGrader.cs:69) and the two UI labels (RoomNarrative.cs:452, SupervisorPhaseSource.cs:188) through AsLine. This moves evidence bytes, so it re-pins.
• ValidateAuthored (AgentAcceptanceContract.cs:250) only rejects an all-blank command, so SupervisorAcceptanceOverlay.cs:49 accepts a blank-first TestsPass amendment such as [" ","dotnet","test"], which Rehydrate.cs:806-815 later nulls and never grades. AgentCodeNode.cs:274 has the same gap. Fix: ask NormalizeOrNull, add the sites to AsksTheHelper, add overlay and rehydrate tests.
• TaskLaunchService.cs:427 (the Delivery/Unattended mandate) only checks Count > 0.
• The amend card's 200-char cap now counts JSON and can cut mid-escape: cut at an element boundary and say how many elements were left out, and show the check's kind on the card.
• Inventory guard: false positives (e.g. HarnessDescriptor.cs:188, NativeLaunchProtocol.cs:45/70) and missed spellings (\u0000, char.MinValue, list patterns).
• PR Apply, clamp or refuse launch controls on every route #2064 carries a sixth copy of the rule (LaunchControlResolver.IsWellFormedFloor); whichever of the two lands second folds the other's copy, and this PR's guard will name it if Apply, clamp or refuse launch controls on every route #2064 lands first.