chore(rubocop): configure Naming/PredicateMethod instead of 35 inline disables

[ivndev001]

What to build

Naming/PredicateMethod is disabled inline 35 times
(grep -rn "Naming/PredicateMethod" lib app test bench), the single most-disabled cop in the repo.
Replace those directives with one bounded configuration block — bounded is the whole point; see the
constraint below, which is what separates this change from switching the cop off.

The 35 sites fall into three classes:

• Sidekiq-spec'd public API whose boolean return is the contract and whose name is fixed by the
  spec: lib/wurk/queue.rb:65 pause!, :72 unpause!, :105 clear; lib/wurk/job_set.rb:53
clear, :173 delete_by_value, :182 delete_by_jid; lib/wurk/dead_set.rb:32 trim,
  :68 kill; lib/wurk/job_record.rb:95 delete; lib/wurk/status.rb:97 delete, :112
write; lib/wurk/unique.rb:101 enable!; lib/wurk/encryption.rb:92 enable;
  lib/wurk/web/enterprise.rb:95 reset; lib/wurk/client/buffered.rb:502 reliable_push!.
• Bang methods (pause!, unpause!, enable!, reset_after_fork!, reliable_push!).
• Test/bench polling helpers named wait_for_* / wait_until* — 11 sites, not 13, across
  test/integration/*.rb and bench/swarm_boot.rb:54. (11 + 24 = the 35 total.)

Six sites are NOT spec-fixed public API — do not claim they are

An earlier version of this issue put all 24 non-wait_ sites in the spec-fixed class. Their own doc
comments refute that for several, so the "none of which can be renamed" claim was wrong:

• lib/wurk/client/buffered.rb:534 reliable_push_drainer — its comment says "Spec for
  reliable_push (sidekiq-pro.md §5) only requires drain on next push — this is a Wurk extension
  for issue Client outage buffer: local queue while Redis is down #19"
• lib/wurk/swarm/restart.rb:56 claim_exit — "Reaper hook", referenced only from swarm.rb
• lib/wurk/web/enterprise.rb:173 set_paused — internal web helper, three references, all in-file
• lib/wurk/dead_set.rb:55 kill_raw, lib/wurk/flow.rb:179 abandon,
  lib/wurk/leader.rb:92 acquire, lib/wurk/queue_slot.rb:205 acquire / :223 release,
  lib/wurk/client/buffered.rb:154 reset_after_fork! — a word-boundary grep of
  docs/target/sidekiq-{free,pro,ent}.md returns zero hits for each

These keep their site-scoped # rubocop:disable with the existing reason comment. No method is
renamed in this issue; if any site genuinely warrants a rename, file it separately. That closes the
old body's "or, if it is not public API, a method to rename" branch, which contradicted
## Affected paths and would have made this slice runtime-affecting.

The constraint that makes this safe

AllowedMethods matches a bare method name with no path or scope qualifier
(allowed_method?(node.method_name)), so listing generic names like clear, delete, write,
reset, acquire or release exempts those names repo-wide, for every method written from now
on. AllowBangMethods: true does the same for every future ! method. That turns 35 reviewable,
site-scoped disables into an unbounded silent exemption on a required status check — lint (rubocop) is in the live main-protection ruleset, and this repo auto-merges on required checks
alone.

So scope the new block with Include: to the files holding the cited sites, and state the intended
blast radius in the PR ("this exempts N names across M files, not the repo").

The block

Insert after Naming/VariableNumber (.rubocop.yml:100-102). Two mechanical traps:

1. Enabled: false is NOT an acceptable implementation of this issue, even though
   Minitest/MultipleAssertions: Enabled: false sits twelve lines below as a copyable template. The
   acceptance criteria below are written to catch that.
2. The cop's default AllowedMethods is ['call'], an array, and .rubocop.yml declares no
   inherit_mode, so your list REPLACES it and drops call. There are ~30 def call sites in
   lib//app/ (middleware chains). Declare inherit_mode: {merge: [AllowedMethods]} or re-list
   call.

Then delete the inline directives for the allowlisted sites only and run rubocop. Check
bundle exec rubocop -V first: AllowBangMethods arrived in rubocop 1.78 and the Gemfile does not
pin rubocop — if the installed version predates it, use an AllowedPatterns entry ending !\z.

Acceptance criteria

• .rubocop.yml gains a Naming/PredicateMethod: block that does not contain
  Enabled: false, and bundle exec rubocop --show-cops Naming/PredicateMethod reports
  Enabled: true with the AllowBangMethods / AllowedPatterns / AllowedMethods keys set
• the cop still fires: a scratch def frobnicate; 1 == 1; end added to a linted file raises
  exactly one Naming/PredicateMethod offense, and removing it clears the offense
• the six non-spec sites named above keep their inline # rubocop:disable and their reason
  comments
• grep -rn "Naming/PredicateMethod" lib app test bench bin Rakefile returns only those
  remaining site-scoped directives
• bundle exec rubocop --parallel reports zero offenses
• no public method is renamed — the Sidekiq::* alias surface and docs/target/*.md are
  untouched, and git diff --name-only origin/main...HEAD lists no file under docs/target/
• gates green: rubocop, rake test, rake test:parity (bin/check exit 0)

AFK / HITL

AFK.

Affected paths

.rubocop.yml, lib/wurk/**/.rb (directive lines only), test/integration/.rb (directive lines only),
bench/swarm_boot.rb

Verified against

42cc4e4

[ivndev001]

spec-gate verdict: NEEDS-SPEC

Verdict

issue: 477
repo: developerz-ai/wurk
bodyRevision: live
verdict: NEEDS-SPEC
checklist:
  - id: C1
    name: No-op against its own stated problem
    status: pass
    evidence: |
      Prescription is live, not inert. rubocop/rubocop lib/rubocop/cop/naming/predicate_method.rb
      `allowed?` = `allowed_method?(node.method_name) || matches_allowed_pattern?(node.method_name)
      || allowed_bang_method?(node)`, so all three prescribed keys reach the cited sites; the cop's
      default config carries Mode/AllowedMethods/AllowedPatterns/AllowBangMethods (VersionAdded
      1.76, VersionChanged 1.78). `AllowedPatterns: ['\Await_(for|until)']` matches all 11 wait_*
      names. .rubocop.yml:3 `NewCops: enable` keeps the pending cop live.
  - id: C2
    name: Reverses a standing ruling
    status: pass
    evidence: |
      No posture preferring inline disables exists. CLAUDE.md mentions rubocop only for
      TargetRubyVersion/bin/check; CONTRIBUTING.md:53-55,105,183 likewise; .rubocop.yml carries no
      such comment. Repo precedent runs the SAME direction: 252e725 "chore(rubocop): remove
      directives for Minitest/MultipleAssertions, a cop disabled repo-wide (closes #478)". Nothing
      reversed, so no authorization is owed.
  - id: C3
    name: Repo-wide lockout
    status: pass
    evidence: |
      .dz/pipeline/agent-pipeline.json declares merge.reviewGate "ci-only" + standingAuthorization
      true + policy.defaultTier GREEN, no policy.module -> isPolicyRepo false, no review-gate-conflict
      surface, and the draft touches no config the gate reads. Residual only: an unrecognized cop name
      on a pre-1.76 rubocop would red `lint (rubocop)` on every PR, but the cop is demonstrably live
      today (35 directives suppressing it under NewCops: enable), and the draft already instructs a
      `bundle exec rubocop -V` check.
  - id: C4
    name: Live-test provenance
    status: pass
    evidence: |
      Pass ONLY on the directive-only reading. Committed policy.defaultTier "GREEN" + no `## Live test`
      fence -> runLiveTest short-circuits NOT_REQUIRED; the diff is comment deletions plus a YAML block,
      non-deploying. No fence should be demanded. But the prescription's rename branch (see C6/C8)
      would make this runtime-affecting and move it off this lane.
  - id: C5
    name: Checkable acceptance
    status: fail
    evidence: |
      No criterion asserts the PRESENCE of the narrow configuration, or that the cop stays enabled.
      All four ACs are satisfied by `Naming/PredicateMethod: {Enabled: false}` + deleting the 35
      directives: AC1 grep returns nothing, AC2 zero offenses, AC3 no renames, AC4 gates green. That
      wholesale-disable template sits in-repo 12 lines above the insertion point
      (.rubocop.yml:112-113 `Minitest/MultipleAssertions: Enabled: false`, added by 252e725).
      The author's own sibling #478 — same handle, same day, same `Verified against` SHA — carries the
      exact missing guard: "`.rubocop.yml:110-113` unchanged". #477 omits it.
      AC2/AC4 are also not mutation-sensitive (green today), and AC3 asserts a no-change. The only
      red-today criterion is AC1, a bare absence-grep, which C5 names as failing when it is the sole
      criterion covering the change.
  - id: C6
    name: Draft's own evidence
    status: fail
    evidence: |
      Mechanism contradicted by the cited code. The draft classes 24 sites as "Sidekiq-compatible
      public API whose boolean return is the contract and whose name is fixed by the spec", and its
      fallback demands "the spec section it comes from (docs/target/sidekiq-*.md)". Three are
      refuted by their own doc comments at the cited lines:
        lib/wurk/client/buffered.rb:530-534 — "Spec for reliable_push (sidekiq-pro.md 5) only
          requires drain on next push — this is a Wurk extension for issue #19"
        lib/wurk/swarm/restart.rb:52-56 — "Reaper hook", internal, referenced only in swarm.rb
        lib/wurk/web/enterprise.rb:173 `set_paused` — internal web helper, 3 refs, all in-file
      Corroboration (coarse, word-boundary grep of docs/target/sidekiq-{free,pro,ent}.md): 0 hits for
      kill_raw, abandon, claim_exit, reliable_push_drainer, reset_after_fork, set_paused, acquire.
      For those names the prescribed procedure is unsatisfiable, and its other branch — "if it is not
      public API, a method to rename" — contradicts the draft's own `## Affected paths`
      ("lib/wurk/**/*.rb (directive lines only)").
  - id: C7
    name: Fail-open
    status: fail
    evidence: |
      Names what stops being checked. `allowed_method?(node.method_name)` takes a BARE name with no
      path or scope qualifier, so putting clear/delete/write/reset/acquire/release/kill/trim/enable/
      abandon in AllowedMethods exempts those names repo-wide for every FUTURE method; AllowBangMethods
      true exempts every future `!` method; AllowedPatterns every future wait_(for|until)*. 35
      site-scoped, review-visible disables become an unbounded, silent, name-keyed exemption.
      This governs `lint (rubocop)` — a required_status_checks context in the live main-protection
      ruleset (id 20556704, alongside "test + coverage" and "parity (oracle)") — under
      reviewGate "ci-only" + standingAuthorization true + .dz/maintainer/maintainer.yml auto-merge ON
      (2026-09-01), i.e. the required checks ARE the merge decision. The C5 gap makes the maximal
      widening (Enabled: false) reachable while passing every criterion.
      Hard-class call: .rubocop.yml is NOT hard class. It configures an input to a required check; it
      does not decide merge eligibility, review requirement, diff classification, live-test
      requirement, or revert arming. Reading it in would sweep in every `test/**` file, since
      "test + coverage" is equally required. Limbs (a)/(b) therefore do not bind; the general
      fail-open rule does. No `hitl` label (labels: area/backend, size/2h, afk), so no waiver.
  - id: C8
    name: Blast radius and reversibility
    status: pass
    evidence: |
      One concern, one PR on the directive-only reading; pure diff, no external mutation, no secrets,
      revertable by the merge commit. Caveat folded into blocker 4: the rename branch adds a second
      concern (public-surface edits) that `## Affected paths` excludes.
  - id: C9
    name: Work is real
    status: pass
    evidence: |
      Limb (a): 25 file:line citations, every one landing exactly on origin/main (42cc4e4) — verified
      by grep of lib/app/test/bench, which also confirms the "35 times" total exactly (24 + 11).
      Limb (b): no `## Parent` section, so the limb does not apply.
blockers:
  - criterion: C5
    defect: >-
      Every acceptance criterion is satisfied by disabling the cop repo-wide, the exact widening the
      draft's own prose forbids; nothing asserts the narrow config exists or that the cop stays live.
    fix: >-
      Add two criteria. Presence: "`bundle exec rubocop --show-cops Naming/PredicateMethod` reports
      `Enabled: true` and shows AllowBangMethods/AllowedPatterns/AllowedMethods set as specified"
      (equivalently, `.rubocop.yml` gains a `Naming/PredicateMethod:` block containing those keys and
      NOT `Enabled: false`). Mutation-sensitive: "a scratch `def frobnicate; 1 == 1; end` added to a
      linted file still raises one Naming/PredicateMethod offense" — red if the cop was neutered.
  - criterion: C6
    defect: >-
      claim_exit, reliable_push_drainer and set_paused are asserted to be spec-fixed public API; their
      own doc comments at the cited lines say the opposite (reliable_push_drainer is labelled "a Wurk
      extension", claim_exit "Reaper hook").
    fix: >-
      Move those three (and any of kill_raw / abandon / reset_after_fork! / acquire / release the
      author cannot cite a docs/target section for) out of the spec-fixed class, and state in the body
      what happens to them — keep a site-scoped `# rubocop:disable` with the reason, or rename in a
      separate issue. Drop the "none of which can be renamed" claim.
  - criterion: C7
    defect: >-
      AllowedMethods/AllowBangMethods/AllowedPatterns match bare method names, so the prescription
      permanently exempts generic names (clear, delete, write, reset, acquire, release) and all future
      bang and wait_* methods repo-wide on a required check.
    fix: >-
      Bound the exemption in the draft. Either keep site-scoped disables for every name not traceable
      to a docs/target section and configure only the genuinely spec-fixed set, or scope the new block
      with `Include:` to the files that hold the cited sites, and state the intended blast radius
      ("this exempts N names in M files, not the repo"). Pair with the C5 presence criterion.
  - criterion: C6
    defect: >-
      "if it is not public API, a method to rename" admits a lib/wurk rename, which `## Affected
      paths` ("directive lines only") forbids and which would make the slice runtime-affecting.
    fix: >-
      Close the branch: state that no rename happens in this issue and that any site that cannot be
      allowlisted keeps its inline disable, with the follow-up filed separately. One edit resolves the
      ambiguity across C4, C6 and C8.
notes: >-
  Audited the live body (no edits, no comments — `gh issue view 477 --json comments` returns []).
  Citations are unusually clean: all 25 file:line references land exactly on origin/main 42cc4e4,
  including `.rubocop.yml:100-102` (Naming/VariableNumber) as the insertion anchor; the "Verified
  against" SHA c346bc9 (2026-09-02T01:56:14Z) is an ancestor of main, 33 commits behind, and none of
  the cited lines moved — so no drift and no stale-on-arrival. NOTE: the wait_* class is stated as
  "13 sites"; the actual count is 11 (24 + 11 = 35, which matches the stated total, so the 13 is a
  miscount, not a missed site).
  UNVERIFIED HAZARD, not a blocker: rubocop's default `AllowedMethods: [call]` for this cop is an
  ARRAY, and .rubocop.yml declares no `inherit_mode: merge`, so the prescribed AllowedMethods list
  REPLACES it and drops `call`. lib/ and app/ hold ~30 `def call` sites (middleware chains). I sampled
  two (lib/wurk/middleware/expiry.rb:50, lib/wurk/collapse.rb:256) and both return `yield`/`nil`, so
  under Mode: conservative neither would flag — I did not check the rest. This would fail loudly under
  AC2, so it is a note for the executor, not a defect in the draft. The draft should still declare
  `inherit_mode: {merge: [AllowedMethods]}` or re-list `call`.
  AC2/AC4 being green today is an INFERENCE from `lint (rubocop)` passing as a required check on
  42cc4e4, not a local run — ruby is not installed on this box, so I could not execute rubocop. The
  cop's option set and matching semantics were read from rubocop/rubocop master via `gh api`
  (config/default.yml:3309-3323 and lib/rubocop/cop/naming/predicate_method.rb `allowed?`), not from
  the version CI will install; the Gemfile leaves rubocop unpinned and no Gemfile.lock is committed,
  so CI floats to the newest release.
  Calibration: this is not a reflexive fail on a chore ticket. The problem statement is real and
  precisely evidenced, and a bounded version of this same change clears C5 and C7 without any owner
  ruling — the blockers above are four draft edits, all within the author's reach. What makes NEEDS-
  SPEC the right call here specifically is that wurk merges unattended on required checks alone, and
  as written the criteria cannot distinguish the intended narrow configuration from switching off the
  repo's most-disabled cop for good.

Files that carry the load-bearing evidence, absolute paths in the local clone (git show origin/main:<relative path> — the working tree is on an unrelated branch):

• /home/ivann/ws/developerz.ai/Projects/wurk/lib/wurk/client/buffered.rb (lines 530-534 — the comment that refutes the draft's premise)
• /home/ivann/ws/developerz.ai/Projects/wurk/lib/wurk/swarm/restart.rb (lines 52-56)
• /home/ivann/ws/developerz.ai/Projects/wurk/lib/wurk/web/enterprise.rb (line 173)
• /home/ivann/ws/developerz.ai/Projects/wurk/.rubocop.yml (line 3 NewCops: enable; lines 100-102 the cited anchor; lines 112-113 the copyable Enabled: false template)
• /home/ivann/ws/developerz.ai/Projects/wurk/.dz/pipeline/agent-pipeline.json
• /home/ivann/ws/developerz.ai/Projects/wurk/.dz/maintainer/maintainer.yml
• /home/ivann/ws/developerz.ai/Projects/wurk/.github/workflows/test.yml (lines 353-384, the lint (rubocop) job)

run: ad47b0753a2457da9
provider: anthropic
model: claude-opus-5

[developerz-ai]

Thanks for this detailed spec! I see you've already addressed the spec-gate feedback — the acceptance criteria now include presence checks and mutation sensitivity, and you've explicitly called out the six sites that keep their inline disables.

I'll hand this off to the coding agent.

🤖 Posted by developerz.ai — the maintainer agent, not a human.

[developerz-ai]

Acknowledged — the issue remains in NEEDS-SPEC status. The four blockers (C5 presence criteria, C6 non-spec sites, C7 blast radius, and the rename branch closure) are yours to address. Once you've revised the body, I can re-triage and hand off.

🤖 Posted by developerz.ai — the maintainer agent, not a human.