fix(GATE-CI-ENFORCEMENT-FLOOR): the diff-scoped walk gets a floor, and its base rule gets a test (#1809)

[localai-bot]

fix(GATE-CI-ENFORCEMENT-FLOOR): the diff-scoped walk gets a floor, and its base rule gets a test

commit-protocol-tag and documentation-checkpoint have been red on main
since 2026-08-13 and cannot recover on their own. Their walk bases on the head
of the last SUCCESSFUL push run, so no green run means the base never advances,
and every later push re-walks the same violations over a range one commit wider.

Re-measured at bacb71109: LAST_GREEN is fafa16f0f of 2026-08-13T01:53, the
range is 499 first-parent commits with no merges, and 41 distinct commits in it
have no available remedy. 35 fail check-commit-trailers.py --range, of which
20 carry no FOLLOWING_AGENTS_PROTOCOL marker at all and the other 15 carry it
one to nine times in a form the strict contract rejects; 6 disjoint ones fail
check-role-discipline.py. check-now-current.py passes over the whole range
and forgives nothing. Every one of the 41 is on main, and the only repair is a
rewrite AGENTS.md forbids.

What changes

scripts/ci-enforcement-floor.txt records one commit the walk never goes
behind. The base becomes the newer of the last green commit and that floor,
decided by git merge-base --is-ancestor rather than by a commit date an author
controls and a rebase can move backwards.

The four byte-similar copies of the base selection in .github/workflows/ci.yml
collapse into scripts/ci-walk-base.py.

The property that survives

A cancelled run stays lossless. The floor is a LOWER CLAMP on an otherwise
unchanged base, so while it sits behind the last green commit the resolved base
is byte-identical to today's. The one window that skips anything is
last_green..floor, which opens at an advance and closes at the next green push
run; it is the forgiveness being asked for, what it skips is fixed at the
advance rather than growing, and the spec enumerates all 41 commits with sha,
date and subject.

That window is currently OPEN and will stay open for a while, which the first
draft of this change described backwards. LAST_GREEN is fafa16f0f and has
not moved since 2026-08-12T23:53Z: of the 100 most recent push runs of
ci.yml on main, 93 are cancelled, 6 failed and none succeeded. So the floor
sits AHEAD of the last green commit, not behind it, the resolved base is the
floor on every push, and the walk grows by one per merge until a push run
finishes green. Correctness is unaffected — floor..HEAD still covers every
post-floor commit — so this is an accuracy repair to the argument, not to the
mechanism.

Proved rather than asserted. A violating commit whose run is cancelled is still
named by the next run's gates, and under the naive github.event.before base
rev-list before..head covers it with nothing, which is #863 exactly.

Why the extraction is the point

Of the four inline blocks, exactly one was ever replayed by a test —
agent-record's, by test_main_baseline.py::AgentRecordDiffRangeTests, and only
under a shim that stubs every python3 call, so it pinned which checker ran with
which range string and could not see the base rule. The other three were executed
by nothing. Nothing in the tree could have caught the ratchet.
tests/scripts/test_ci_walk_base.py is 31 cases over a real throwaway
repository, because every question here is about ancestry.

What is narrowed, and why

41 commits stop being walked. Enforcing on an immutable landed commit is not
enforcement: no contributor action clears it, so it is a permanent red rather
than a gate, and a permanent red is read by nobody — which is #1722's effect and
strictly worse than a smaller gate that is read. The gate's purpose is to stop a
NEW violation and that is untouched. No assertion is deleted, no checker is
modified, and advancing the floor stays a reviewed commit that must name what it
forgives.

A per-commit exemption list was evaluated and rejected: AGENTS.md has no waiver
registry, the mechanism would have to be built three more times in checkers that
have no exemption concept, and it never shrinks the walk, so the range would
grow by one per merge forever. Reverting to github.event.before was rejected
because it reintroduces #863 outright.

Evidence

Measured by reading the gates' run: bodies out of ci.yml with a YAML parser
and executing those exact bytes, so nothing about the gates is transcribed.

Case	grep	strict	doc-checkpoint
the implementation commit, floor in place	0	0	0
the floor moved back to the frozen base	1	1	1
the four resolver call sites deleted	1	1	1
a new commit with no marker	1	1	0
a new product commit naming no pull request	0	0	1
the control: same path and trailers, (#1809)	0	0	0
the floor record emptied to a comment	2	2	2

The last row is fail-closed, and it is a measurement rather than an intention: a
broken floor record read as "no floor" would silently restore the ratchet, so
the parse refuses and every step aborts under set -eu before invoking a single
checker.

The call-site deletion also reds test_ci_walk_base.py at 2 of 31. Both
mutations were restored and each restoration verified by sha256sum -c with a
clean git status.

One trap is recorded in the spec rather than quietly fixed: the first
no-trailer mutation named the marker in its own body while explaining that it
lacked one, which satisfied the presence grep and read as a passing gate.

Three existing assertions that matched the old inline shell as a string moved
rather than went away. test_main_baseline's shim now executes the resolver
instead of stubbing it, and its before fallback case and test_agent_gates's
role-range case now EXECUTE the resolver instead of grepping the YAML — stronger
than the literals they replace.

A fresh review returned FAIL, and this is the repair

Five findings, all in the RECORD. The mechanism passed: scripts/ci-walk-base.py,
the recorded floor sha, .github/workflows/ci.yml, tests/scripts/test_ci_walk_base.py
and the three checkers are byte-identical, and git diff --stat against the
reviewed head touches one file under .agents/.

• Two of the 41 enumerated shas resolved to nothing. ddff09093663 and
  be4a3edf1727 are not objects; the commits are ddff090936bb and
  be4a3edf17b2. That is load-bearing rather than cosmetic, because
  scripts/ci-enforcement-floor.txt says of that table "That record is the only
  place that now says those violations happened". All three sites are corrected
  and the whole table is now verified rather than spot-checked: 41 of 41 unique
  shas resolve under git rev-parse --verify -q '<sha>^{commit}', and every
  recorded date agrees with the commit's author or committer date.
• A mitigation was claimed that does not exist. Risk 2 said a too-far-forward
  floor is caught by ancestry and by RecordedFloorTests. Those guards catch an
  OFF-HISTORY floor. A floor advanced too far to a commit that really is on
  main IS an ancestor of HEAD, so both accept it — measured with the floor
  set to HEAD itself, where known and is_ancestor are both true, the
  resolver returns HEAD, and all three diff-scoped steps return rc 0 vacuously
  over an empty walk. The risk now says the only mitigation is the reviewed
  advance. This is a different failure from the fail-closed case in the evidence
  table above, and both are stated as what they are.
• The floor swallows the one in-checker exception it cites as precedent.
  281b4bc76c0e (A landed commit on main carries a malformed Assisted-by, because the PR BODY was never corrected with its commits #1262) is below the floor, so
  check-commit-trailers.py --range fafa16f0f..origin/main prints its
  This is DEBT, not success banner once and --range bacb71109..origin/main
  prints it zero times. The entry is not deleted or bypassed and
  LandedMessageExceptions still pins it; only its runtime visibility on the
  push lane is lost, and the spec now says so.
• The steady-state claim was wrong, corrected above and in two places in the
  spec.
• The floor forgives by RANGE, not by violation. A checker written tomorrow
  that finds a new class of defect in the pre-floor range is forgiven silently
  too, which an exemption list would not do. That is the strongest argument
  against the shape chosen here and the comparison was one-sided without it.

The floor is advanced once, and this is the first exercise of that mechanism

Risk 1 arrived one day after the floor was recorded, and it is now
discharged. Re-measured independently at origin/main d60692c89, over
bacb71109..origin/main, ten first-parent commits: the presence grep is rc 0,
check-commit-trailers.py --range is rc 0, check-now-current.py is rc 0, and
check-role-discipline.py is rc 1 on exactly one commit.

That commit is c00b99c7c, fix(LTX25-DIT-ATTN-ARM-PARSE): match every DiT attention arm exactly and refuse a fourth value, 2026-08-24. It landed
src/vllm/model_executor/models/ltx2_device.cpp and
tests/vllm/models/test_ltx2_device.cpp — product code and its own test —
straight onto main with no row/<ID> branch in its history. Its message is
clean and passes the trailer contract, so this is a role-discipline
violation and not a trailer one: the split across the forgiven set moves from 35
trailers / 6 role discipline to 35 / 7, and the count from 41 to 42.

Why the advance is justified. c00b99c7c is on main and can only be
repaired by rewriting main, which AGENTS.md forbids. It is the same condition
as the 41 already enumerated and it produces the same failure at a smaller
scale: with the floor left at bacb71109, the agent-record role-discipline
step would red on that one commit on every future push, forever, with no action
any contributor could take to clear it. That is this row's own defect arriving
at its own merge — one commit instead of 41, and freezing again on the next one.

Why this far and no further. The floor is set to c00b99c7c itself, not to
origin/main. The walk is FLOOR..HEAD and excludes FLOOR, so this forgives
that one commit and leaves e6f4f566f, d60692c89 and everything after them
enforced. Advancing to current main would have been the same edit and would
have forgiven two commits nobody examined. Risk 2 records that nothing in the
mechanism can detect that; the review is the only control, and minimality is how
it is honoured.

c00b99c7c is enumerated as the 42nd forgiven commit with its sha, date,
subject and what it did. All 42 shas were verified to resolve —
git rev-parse --verify -q '<sha>^{commit}' over all 42: 42 resolved, 0
missing, all 42 ancestors of origin/main. A full sweep, not a spot-check,
because after this lands the enumeration is the only witness that these
violations happened, and an earlier round of this list carried two shas that
resolved to nothing.

Measured after the edit, each rc captured on the command itself and never after
a pipe:

Floor	grep	strict trailers	role discipline	check-now-current.py
bacb71109 (old)	0	0	1, c00b99c7c	0
c00b99c7c (new)	0	0	0	0

git merge-base --is-ancestor c00b99c7c origin/main is rc 0, and
tests/scripts/test_ci_walk_base.py is 31 tests, 0 failures over the new value.
The gate still bites: a scratch commit above origin/main carrying no
FOLLOWING_AGENTS_PROTOCOL string anywhere — marker count printed as 0
before the run, because a mutation whose own message mentions the marker
satisfies the presence grep and reads as a pass — reds both trailer steps by
name over c00b99c7c..<scratch> while role discipline and check-now-current.py
stay rc 0. It was built with git commit-tree and never entered the worktree;
sha256sum -c on both changed files passes afterwards.

Two things follow from this being the first advance, and they point in opposite
directions. The mechanism works, and it was exercised before the pull request
introducing it had merged. And the violations are still arriving: c00b99c7c is
the second role-discipline violation in two days, after 6e73bdee3ea1. A floor
stops an unrepairable commit freezing a gate. It is not a fix for whatever is
putting product code on main without a task branch.

main moves roughly every twenty minutes, so whoever merges repeats the four
commands over c00b99c7c..origin/main and, if one reds again, advances the
floor the same way.

That re-measurement has been done once more, at the merged head. main
moved to df1ee2058, four first-parent non-merge commits above the floor. Each
rc captured on the command itself and never after a pipe:

Gate	Command	rc
commit-protocol-tag presence grep	the loop replayed out of ci.yml, not transcribed, over c00b99c7c..origin/main	0
strict trailers	python3 scripts/check-commit-trailers.py --range c00b99c7c..origin/main	0
role discipline	python3 scripts/check-role-discipline.py --base c00b99c7c --head origin/main	0
now-current	python3 scripts/check-now-current.py	0

git merge-base --is-ancestor c00b99c7c origin/main is rc 0, and
tests/scripts/test_ci_walk_base.py is 31 tests, 0 failures, rc 0 at the merged
head. The recorded floor is still sufficient and still minimal, so nothing
about it changes here: scripts/ci-enforcement-floor.txt is byte-identical to
the reviewed head and still reads
c00b99c7c8b64f9247230ed6220598cc5c0e347e. Whoever merges repeats the four
commands one more time, because main keeps moving.

The one review finding, repaired

A fresh review returned PASS with one nit, and this is it.
.agents/specs/ci-enforcement-floor.md presented check-role-discipline.py's
error inside a fenced block as if it were verbatim, but the block was re-wrapped
and truncated after …reached main without arriving on a task branch. with no
ellipsis marking the cut. The dropped half is the half that states the rule —
work happens in its own worktree on a row/<ID> branch and lands through a
reviewed PR or an authorized local merge naming that branch — which is the
reason c00b99c7c is a violation at all. A record that argues for forgiving a
commit was doing it with the weaker half of the evidence.

The full message is pasted rather than an ellipsis appended, because the fenced
block is part of what justifies the advance. It is the checker's single
unwrapped line, copied from its own stdout under
--base b207f34d3 --head c00b99c7c (which exits 1) and compared byte-for-byte
against the file, not retyped. The prose now names that invocation so the next
reader reproduces it instead of trusting a transcription.

git diff --stat against the reviewed head 224f9afb6 shows this commit
touching exactly one file, .agents/specs/ci-enforcement-floor.md, 5 insertions
and 4 deletions. scripts/ci-walk-base.py, scripts/ci-enforcement-floor.txt,
tests/scripts/test_ci_walk_base.py and the three checkers are untouched, and
.github/workflows/ci.yml changes only by what the merge of df1ee2058 brings
in.

Merge-time hazards, recorded rather than fixed here

• The first origin/main conflict was real and is resolved; the second
  CONFLICTING was a GitHub artefact. The trial merge against af320abb2
  genuinely conflicted in scripts/agent-preflight.sh's SUITES array, so
  GitHub could produce no merge ref; 3574065e7 is merged in and both sides'
  entries are kept, af320abb2's three plus this branch's test_ci_walk_base.
  main then moved again and GitHub reported CONFLICTING/DIRTY a second
  time, for a merge Git performs cleanly. The only file both sides touched was
  .agents/issue-index.md, which .gitattributes:7 marks merge=union —
  an attribute GitHub does not apply when it computes mergeability. df1ee2058
  is now merged in as well, with Auto-merging on that file and on ci.yml and
  no conflict anywhere. Merging is also what makes the advanced floor
  recordable, since RecordedFloorTests::test_recorded_floor_is_an_ancestor_of_head
  asserts the recorded value is an ancestor of HEAD and c00b99c7c landed
  after this branch left main.
• The .agents/issue-index.md row survived both merges exactly once, and the
  second merge was verified by hand rather than trusted to the union driver,
  which has twice silently dropped a tail row in this repository. Merge base
  3574065e7 carries 659 rows; this branch adds #1809, main adds #1823,
  and neither side removes one. The merged file carries 661 rows whose set is
  exactly the union of the two sides — compared as sorted sets, not counted —
  with #1809 and #1823 each appearing once, no duplicate row, the preamble
  byte-identical to both parents, and zero conflict markers across all six files
  the merge touches. The row is no longer the tail — later rows follow it — which
  is correct for an append-only union file, and it is not edited: the row's own
  prose still says 41, because an index row is never rewritten.
• scripts/agent-preflight.sh has no SKIP left, and that is now measured at
  the merged head rather than at the previous one. commit-trailers and
  commit-style SKIPPED whenever origin/main was not an ancestor of HEAD,
  and the script prints NOT a green preflight: a skipped gate reported nothing about this tree itself when that happens — a rc 0 carrying a SKIP is a third
  state the exit status cannot express. Rerun with a role claimed in the working
  worktree after merging df1ee2058: zero SKIPS, the string does not appear
  in the log, NOT a green preflight does not appear, and both gates report
  ok under Commit trailers vs origin/main df1ee20584d69d2ac434c7672f9099c71abc7f49,
  alongside ok now-current range and ok issue-index append-only. The spec's
  ## Outcome line claiming "All gates green" was already corrected to say
  this.
• scripts/ci-walk-base.py's docstring still carries the corrected
  steady-state sentence. It was left alone so the reviewed mechanism stays
  byte-identical; it is named in the spec so it is not left to be found.

test_cpu_x86_llamacpp_floor is red in preflight and is #618, not this row, so
the run exits rc 1 on that one gate and nothing else. It was discriminated
rather than asserted: pristine origin/main at 0a0a53e5a, run serially in its
own clone, fails the same CpuX86FloorHarnessTests cases with
NO_QUIET_WINDOW at loadavg 63-68. The rerun at the merged head reproduces
#618's recorded signature exactly —
test_a_contended_leg_is_discarded_and_never_summarised asserting
got.returncode == 2 and getting 4, waiting for quiet: 15s busy=112% builders=0 load=36.77 — which is the NO_QUIET_WINDOW (4) for GIVING_UP (2)
substitution the index row for #618 describes. The box carried loadavg 38-52
throughout, with a second session's preflight running concurrently.

Closes #1809.

FOLLOWING_AGENTS_PROTOCOL

Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: AGENT:claude-opus-5 [Claude Code]