revert: bound serialized writer contention

[brokemac79]

Summary

Revert #727 in full and restore the exact publication controller and state-lease timeout from immediately before that merge.

Why this rollback

#727 deliberately clamped state-contention publication to four workers. The control activated as designed, but the immediate production samples moved in the wrong direction for the incident:

• pending: 498 -> 518 -> 531 -> 536 -> 539
• latest 15-minute net drain: -12/hour
• latest useful delivery: published 20/hour + superseded 4/hour versus arrivals of 52/hour

The first fully constrained cohort did produce one successful publication_applied result, so this rollback is an operator-directed emergency decision rather than proof that every #727 mechanism was incorrect.

Scope

• exact Git revert of merge commit 74ccb2ecab50916f7a4372b823dac25d1c7d11f3
• restores the complete tree from pre-fix(publish): bound serialized writer contention #727 main commit 6444e678dccb3cce1505827674b05a1f2f7ce397
• no changes to fix(publish): serialize exact state mutations #722/fix(publish): recover exact state races without full reset #723/fix(publish): coordinate every state branch writer #724 state-writer serialization and fencing
• no queue replay/reset, dead-letter replay, gate change, workflow dispatch, or retry mutation

Validation

pnpm run build:repair
pnpm run build:dashboard
node --test test/dashboard-worker.test.ts
  156 passed, 0 failed
node --test --test-name-pattern "state publish lease" test/repair/git-publish.test.ts
  5 passed, 0 failed
pnpm run lint:repair
pnpm run lint:dashboard
pnpm run lint:scripts
git diff --cached --check
git write-tree == 6444e678dccb3cce1505827674b05a1f2f7ce397^{tree}

The Windows checkout formatter reported the established line-ending noise on the four reverted files. The staged tree was independently verified to equal the already-green pre-#727 tree exactly.

At the maintainer's explicit emergency direction, the normal additional Codex-review and Crabbox rerun gates were waived for this exact rollback. The focused tests above had already completed successfully.

Rollback risk

This restores the pre-#727 behavior:

• adaptive publication can again admit 24–48 publishers;
• state-lease acquisition returns to 180 seconds;
• the prior high-contention retry/dead-letter storm may return while all publishers serialize behind the one state writer.

Post-merge monitoring must compare successful published + superseded delivery with arrivals and must not mistake dead-letter disposal for recovery.

Attribution

This reverts the narrow CSW-049 follow-up authored in #727. It leaves the earlier state-safety work and contributor attribution in #722, #723, and #724 intact.

No automerge is requested.