Consume the dead-code gate from ops instead of vendoring it

[jonnyspicer]

Part of 86cb417ty. Same change as the merged pilot, retina-analytics#13.

Why

tools/check-dead-code.sh reported success when vulture was not on PATH — the shell's command-not-found went to /dev/null and || true discarded exit 127, so the gate printed no dead code found and exited 0. It only ever produced false passes, so nothing surfaced it until CI disagreed with a local run.

The header also claimed a canonical copy lived in offworldlabs/ops. It never did: the file was vendored byte-identically in six repos with no upstream, so there was no single place to fix it.

What changed

ops is now a pre-commit hook repository, and this repo pins it:

repos:
  - repo: https://github.com/offworldlabs/ops
    rev: dead-code-v1.0
    hooks:
      - id: dead-code

The vendored copy is deleted. The rev: pin is the source of truth, so drift is impossible rather than merely detectable — the mechanism claude-shared's scaffold already uses for ruff. Future fixes reach every repo by bumping one line; pre-commit autoupdate does that for you.

The gate now fails closed: absent vulture exits 127 naming the tool.

Not in scope

ruff stays a separate CI step. Aligning these repos with claude-shared's ci-python.yml is worth doing but belongs in its own ticket, not a bug fix.

Verification

pre-commit run --all-files → dead code (vulture)....Passed, exit 0, no findings. Upstream ops CI runs the gate's 9-test suite plus pre-commit try-repo, with REQUIRE_VULTURE=1 so a broken vulture install fails the job instead of silently skipping.

🤖 Generated with Claude Code

https://claude.ai/code/session_013ZcazeseXVpu8XrYA2tE1V