docs(gates): inventory all 81 gates/tiers/batteries/hooks with evidence — 67 have a named catch, 0 earn a DROP

[ZacxDev]

Closes clawgate task #525.

An inventory of all 81 gates, tiers, batteries and hooks devrc runs. Every row carries
either named evidence of a real catch — a session id, a CI run, or a commit sha — or the
explicit string no evidence found with the search that produced it, plus a KEEP/DROP/TIER
verdict.

🔴 Nothing that runs is changed by this PR. No gate, target, battery or hook is dropped,
disabled or made conditional, and the per-path selection rule in §10 is proposed, not
implemented. Both are the task's explicit non-goals; the verdict column is a recommendation
the operator decides on.

How the enumeration was derived

From the repo at scan time, not from the task's own list (which the document then contradicts in
two places — see below). Commands and counts are in §1 so they can be re-run and disagreed with:
31 pytest targets, 5 node suites, 2 nix check derivations, 10 mutation batteries, 16
claude-hooks modules, 13 drift-check arms, the githooks pre-push gate, 2 Tekton checks.

Method, and why the numbers are worth anything

Four transcript scanners over 5,836 .jsonl (6.8 GB), each with a negative and a positive
control, all reported in §2. This session's own transcript is excluded from every count.

🔴 One control FAILED, and the table was not trusted until it was explained. scan_drift.py's
rc=0 positive control returned 0 hits. Cause read out of the source rather than assumed:
scripts/drift-check.sh:3554 guards the (rc=$rc) line behind if [ "$rc" != 0 ], so a clean
run is structurally incapable of emitting that format — it prints
drift-check: no drift on the host(s) CHECKED instead. Re-run against that sentence: 34
sessions. Control passes.

The toolDenialKind co-requirement was load-bearing, and here is the measurement that shows
it: on the same search string, text-only finds 97 sessions and text-plus-marker finds
75. Those 22 sessions discussed a guard without being blocked by it, and would have
inflated every hook row.

What the evidence shows

• 67 of 81 rows carry a named catch. All 10 mutation batteries; 11 of 14 bash-guard arms;
  11 of 13 drift-check arms; 28 of 31 pytest targets red at least once; 4 of 5 node suites; 8 of
  8 hooks that are actually guards.
• 14 read no evidence found, each with its search recorded, and none on idleness
  grounds: six are modules that are not gates and emit nothing to find; two are drift-check
  arms structurally unable to fire yet; three are never-red targets covering guards that have
  fired; three have the evidence in a sibling tier.
• 0 rows earn a DROP.

The clearest case for keep-despite-no-evidence is §7's three zero-firing bash-guard arms
(mkfs, dd to a block device, rm -r $HOME): their cost is a regex on a string already being
matched for eleven other arms, and the loss they prevent is total and unrecoverable. Frequency
is the wrong axis for a guard whose downside is unbounded — which is the same reason a
run-count could never have answered this question.

Three findings, none of them a drop

1. 🔴 scripts/devhost-tests is in TARGET_FLOORS (31 entries) and not in HERMETIC_TARGETS
   (30). It appears in 4 sessions where its neighbours appear in 200–590, while its floor
   still counts toward the derived global floor. This is already the failure mode the task's
   non-goals warn about — a target that silently stopped running — arrived at by omission rather
   than by anyone choosing it. Decide it explicitly, either way.
2. 🔴 No mutation battery is wired into any gate.
   grep -rn "mutants-" scripts/gate.sh scripts/run-tests.sh scripts/run-node-tests.sh flake.nix githooks/
   returns nothing. Defensible as manual one-shot instruments, but currently undocumented — worth
   a line in each battery's header so the next reader does not assume they run.
3. 🔴 The proposed tiering saves less than it looks, and the saving is unmeasured. The
   always-run set is 14,053 of the 21,202 pytest floor-sum (66%), capping the benefit at
   ~34% and only on single-subsystem diffs. Floors are not wall-clock. §11 also lists two risks
   worth answering before building it: a path rule is itself a guard that can fail silently, and
   conditional selection interacts badly with the derived global floor.

Also in this PR

scripts/check-gate-inventory.py asserts criteria 3 and 4 mechanically over the table —
81 rows checked, 0 problems. It carries its own controls (--self-test: four malformed row
shapes each rejected, a well-formed row accepted, plus a scoping control), and it errors rather
than passing when it matches zero rows, because a vacuous green there would certify an empty
table. Its first run reported 11 false positives against §10's rule table, which reuses row ids;
fixed by scoping to the inventory sections rather than by loosening the check.

It is deliberately not wired into any gate and not under scripts/tests/ — a file there
would change that target's collected count and force a TARGET_FLOORS re-pin, which this
document does not need.

Follow-on

Clawgate task #528 — add emit_invocation to the gates and the five blocking guard hooks,
recording the outcome (and, for bash-guard, which arm), so this is answerable from
activity.events directly next time instead of by transcript proxy. An event recording only
"the gate ran" would reproduce the exact gap this document had to work around, so #528 makes the
outcome field a criterion and requires the emitter to be fail-open and demonstrated so.
(#526 was already taken by the session that filed #525.)

Note on criterion 1

The task says "a PR against ZacxDev/devrc". This repo's origin is
innovation-upstream/devrc (git remote -v). Read as "the devrc repo" and targeted at origin.
Flagged rather than silently resolved.

🤖 Generated with Claude Code

https://claude.ai/code/session_01HXxTANDijZTVxj398NrWBU

[ZacxDev]

🔴 Correcting this PR's description — round 1's headline finding is RETRACTED

Posting rather than silently editing the body, because a reviewer may already have read the
wrong version.

The PR body's finding #1 — that scripts/devhost-tests had "silently stopped running,
arrived at by omission" — is false on all three of its claims, and the remedy it proposed
would have made things worse:

body claim	measured
"arrived by omission rather than by anyone choosing it"	scripts/run-tests.sh:876-898 is a dedicated DEVHOST_TARGETS array under a 🔴 comment explaining the choice — it needs a real nvim, added 2026-08-29
"a target that has silently stopped running"	scripts/tests/test_nvim_clipboard_osc52.py parses that array and asserts the registration exists
"its floor of 6 is counted into the derived global floor regardless"	False. run-tests.sh:900 sets TARGETS=("${HERMETIC_TARGETS[@]}") and the floor sums over $TARGETS. The PR's own gate output said floor: 21281 = sum of **30** per-target floors against 31 targets — the refuting evidence was in hand and I misread it

🔴 And "add it to HERMETIC_TARGETS" would have redded the sandbox tier — these tests need
nvim, which the nix sandbox does not carry, which is exactly why they were moved out
(ERROR: 3 UNPINNED skip group(s)). No action is needed; the design is correct as it stands.

Second retraction: the bash-guard arm table was wrong in both directions. It was assembled
from the guard's deny messages instead of from guard_core._CLAUDE_CODE_CHECKS, the list
bash-guard.py:60 actually runs:

• check_git_commit_to_main was missing entirely — re-measured independently of the audit at
  435 blocks across 270 sessions, the third most-fired arm in the repo, and the one
  enforcing this repo's own 🔴 never-commit-to-main rule.
• rm -r $HOME is not a bash-guard arm at all — it is in _IRREVERSIBLE_CHECKS, an
  opencode-only policy, under a 20-line comment arguing the exclusion. It has fired 3 blocks / 1
  session, from opencode's guard.

That one mattered beyond bookkeeping: the body's keep-without-evidence argument read "a regex on
a string already being matched for eleven other arms". For rm -r $HOME under claude-code
that is false — it is not matched at all, so its zero was a policy fact, not a rarity fact.
An operator persuaded by that argument might have moved check_rm_rf_critical into the
claude-code policy, reversing a deliberate and argued decision. Corrected: 12 of 14 arms have
fired; the zero-firing set is two, not three.

Also corrected in round 2, each measured rather than reasoned:

• Row C3 counted readers, not firings. gate.sh:250 emits GATE: RESULT=UNVOUCHED; it emits
  could-not-vouch nowhere — that string lives in CLAUDE.md, which carries both of round 1's
  needles in one sentence. Re-measured: 18 sessions, not 241. The KEEP survives; the figure
  overstated ~13×. A fabricated-name negative control structurally cannot catch this, and that
  limit is now stated in §2.
• The documented search needle was wrong — the runners emit two spaces
  (FAIL <target> (). Round 1's regex used \s+ so its counts were right, but a reader
  following the appendix would have got 0 for nearly every row and read that as confirmation.
• Instrument 3 cannot separate the two tiers — both runners emit an identical
  PASS <dir> ( and three dirs are both a pytest target and a node suite. Six rows are now
  marked AMBIGUOUS, and C2's reason no longer leans on a separation the instrument cannot make.
• §10's always-run triggers omitted scripts/testlib/** (122 tracked files, 8+ targets) —
  a live instance of the risk §11 itself states. Added.
• The base moved twice under this PR. chore(subsystems): retire the initiatives board and repo-cos — skills, code, units #1377 retired initiatives and repo-cos, so round 1
  carried rows for two deleted targets; then two new mutation batteries landed. Re-anchored
  112a5225 → 01956bf0: 81 rows (was 81 → 79 → 81 for different reasons), pytest
  floor-sum 21,202 → 20,348.

The checker had three defects of its own, all now mutation-verified

• Its scoping control was vacuous. The fixture contained no ## 3. line, so in_scope was
  False from the start regardless of the branch it claimed to cover; deleting that branch left
  the self-test fully green. Fixed, and mutation-checked: with the branch deleted the control
  now fails with its own error, so it is reachable.
• Its evidence check accepted vocabulary, not evidence. The bare word CONFIRMED passed.
  Decisive control: replacing all 81 evidence cells with CONFIRMED yielded RESULT: PASS. It
  now requires a lookup-able identifier; the same control fails all 81 rows.
• Its zero-row guard was narrower than its docstring — a document truncated to one row
  passed. The document now declares its size and its verdict tally in machine-read markers; a
  mismatch or missing marker exits 2.

🔴 The tally marker immediately earned its keep: the hand-written tally was wrong again this
round (47/32 written, 49/30 counted), as it was in round 1 (51/30 written, 49/32 counted) —
while the row lists beside it were correct both times. It is now machine-checked rather than
proofread.

The pattern, which is the part worth keeping

Both 🔴 were a set inferred from its members' behaviour instead of read from the registry that
defines it — the arm table from deny messages rather than _CLAUDE_CODE_CHECKS, row A29 from
two array names without reading the 40 lines of comment above them. Every remaining error was a
variant: a needle inferred from prose rather than the emitting echo, a tier separation assumed
rather than checked against the runners' output.

What held up was the mechanical arithmetic — all floors, both floor-sums, 12/12 battery shas,
11/13 drift arms, every quoted line number. The controls did their job where they were pointed;
what they could not do was tell the method it was pointed at the wrong thing. That is recorded
in §13 of the document rather than left in this thread.