Delete the derived-generator-input arm: #10856 fixed the root, so it classifies a delta that no longer occurs

[gunbai-bot]

DO NOT MERGE WITHOUT THE OPERATOR. This reverts a merge the operator made. It is opened so the finding and its evidence survive in a durable place; landing it is theirs.

What this deletes

#10871 (d2559c36e3) added DerivedGeneratorInputResolution: an auto-admission for a 0 -> 1 binding on the derived failure-mode roster whose every candidate is a module of the generator's declared input directory.

That delta no longer occurs, and it was already gone when the arm merged.

The root was fixed first

#10856 (aba7939b54, 2026-09-08T18:51:35-04:00) fixed the cause. Its own doc block states the mechanism:

roster.dag is gitignored and written on the read path, so the diff can never name it — and carrying it made the HEAD's roster stand as the base's … the roster's base and head source, which were THE SAME BYTES, so an ordinary append read as not locally authored and classified NewPoolCoincidenceResolution.

With the base side derived from the base tree's row membership, an ordinary append produces no binding delta at all.

#10871 merged at 23:19:18 — four and a half hours later — classifying the symptom.

Measured

Three floor runs, two branches, all on base 34800395b1:

	DerivedGeneratorInputResolution	NewPoolCoincidenceResolution	deltas
#10842 abf30e7fb0	0	0	2, membership only, ADMITTED
#10842 631cd3bbf0	0	0	2, membership only, ADMITTED
#10710 1ce4fe1eac	0	0	2, membership only, ADMITTED

deltas went 4 → 2 on #10710. There are no binding deltas at all.

A fourth run, on a base that still HAS the arm

#10892 appends one failure-mode row — modules_added=1, exactly the change this arm was built to admit — against base fc464baf92, which contains the arm:

namespace-wave-admission base=fc464baf92 … modules_added=1
ExplicitlyEvaluatedZeroDelta membership …roster -> …mechanism_outlives_the_population_it_was_built_for
ADMITTED — every delta is auto-admitted or named by a transition admission
DerivedGeneratorInputResolution: 0

The arm is present and inert on the precise change it exists for. This is the discriminating observation the three earlier runs only implied: they showed the delta absent, this shows the arm loaded and silent while the append it was written for sails through on membership alone.

Three DESIGN readings, same direction

• §3 — with the root fix landed, the symptom classifier is a second structure answering one question, and the surviving one is an attractor: nearby questions keep getting answered in its vocabulary.
• §3c — its only consumers were the two fixtures shipped with it. Dangling.
• §5 — a populationless arm that widens what a required gate auto-admits can never go red, and will be cited as coverage of a class it never sees.

The deletion is the census. If a real delta of that shape exists, the gate refuses and names it — the evidence the arm never had.

Why no regression probe is added

The fair worry is that the root fix could regress with nothing to catch it. #10856 already enrolled that evidence: the_base_side_is_the_base_listing_not_the_head_directory and a_base_tree_with_no_row_files_has_no_roster_module_rather_than_an_empty_one, both passing on this branch. A second probe would fork existing evidence (§3).

What is true — and is a pre-existing declared drop, not this change's to repair — is that those are #[cfg(test)] under repo_self_test_command, which no CI step runs (gunbc.rung_drop rust_unit_tests_off_the_merge_path). The evidence exists at rung 2 and does not execute on the merge path. That gap is identical with or without this revert. The arm did not close it, because an auto-admission cannot be evidence of anything.

How this got in

The arm was reviewed four times and approved. Every reviewer — and the manager, and me — reasoned from CI evidence bound to base 788013c849, which predates #10856. The review process did not fail to check whether the arm was correct. It failed to ask whether the arm was still needed.

Mechanics

Exact inverse of d2559c36e3 via git revert; nothing has touched these files since it landed, so this is not a hand-reconstruction. 50 wall integration tests pass; #10856's own probes pass.

Correction (review 62904). The first push was not the exact inverse it claimed: git add -A swept in src/v2/test/claim/modules_under_container_test.dag, untracked residue from the closed #10836, whose subject modules_under_container is defined nowhere in the tree. Removed. The claim was the worse half of the defect — "exact inverse" was checkable and false, and a reader who trusted it would not have looked, which is the same shape as the arm this branch deletes. It is now verifiable:

git diff --stat main..HEAD                      # six files, only the arm
git diff d2559c36e3^ HEAD -- <the six files>    # empty: byte-identical to pre-arm

🤖 Generated with Claude Code

https://claude.ai/code/session_01GwvzQ3SQoWWDgHkAeffMeF