v7.16.1

A patch of gate defects found by exercising 7.16.0's own plugin work.
Engine identity unchanged at engine-service-v0.1.85.

Seven fixes: one user-visible regression, three gates that could not
fail, one gate measuring the wrong file, one test coupled to ambient
developer state, and nine eval graders blind to their own subject. Four
shipped in 7.16.0 itself (search, the cut assert, the reconcile tests,
the eval graders); three are older (the lint leg, the watchdog, the
routing cap). They share a shape: the test sat one layer below where
the behaviour lived, so it passed while the thing it named was broken.

Fixed

• search over MCP returned hashes instead of results. The wire
  wrapper added in 7.16.0 attaches structuredContent alongside the text
  block. A conforming client may render structuredContent INSTEAD of
  the text -- the protocol permits it, nexus's own client does it
  (mcp_client/core.py::_parse_result prefers it), and Claude Code does
  too. So a default-mode search delivered ids, tumblers, distances and
  chunk hashes with no file paths, line numbers, snippets or pagination
  footer. The payload now carries the render, so either representation is
  self-sufficient. The server was never wrong and neither were its ten
  tests: every assertion is on the CallToolResult, where both shapes
  are present and correct, and the displacement is not observable from
  there. An AST census now requires any tool building a
  CallToolResult with structuredContent to carry a text key, which
  binds query/nx_answer/memory_search before they are written.
  (nexus-6jlki)
• The plugin-cut ledger assert failed every cut. plugin-release.yml
  grepped each touched path against the whole of PENDING_RELEASE.md;
  the ledger's permanent header names .claude-plugin/marketplace.json,
  and every cut touches that file by construction, so the offender list
  was never empty. The decision moved to
  scripts/check_cut_ledger_clean.py, which scans bullet ENTRIES via the
  existing path_entries() parser. Blocked the first real cut
  (nexus-a2wmi.12). (nexus-a2wmi.9)
• The -m lint CI leg could not fail. pytest ... | tee under
  GitHub's default bash -e (no workflow declares shell:) exits with
  tee's status. Since addopts deselects lint everywhere else, that
  job was the sole enforcement point for ~1050 tests feeding the required
  pytest-gate check -- including the guards against a new MCP tool
  re-opening the regression above. A new lint requires every piping
  run: block in every workflow to set pipefail.
• The scheduled-failure watchdog reported green over 8 red nights. It
  sampled the newest run of any status, and fires 6 minutes after a
  nightly that takes 12-25, so the sample was an in-flight run whose
  conclusion is null -- not a failure, not stale. It now requests
  status=completed. Separately, the Sweep step forwarded any exit code
  but 2 as a verdict, so 127/124/137 closed the tracking issue with an
  all-clear. Verified against the live API: the same command that
  reported 0 findings now names the failing gate.
• The routing-hook cap measured the documentation. It summed
  registry.yaml rules, but hook.py:271 says hooks.json is the
  registration surface -- 3 hooks fire on PreToolUse:Bash against 2
  listed rules, so a fifth could land under a reported 2/4. The RDR-121
  budget is now enforced against the registration surface.
• Stop-hook reconciliation tests failed on developer machines. They
  faked HOME, which drops ~/.config/git/ignore, so the standing
  untracked .claude/settings.local.json resurfaced and tripped the
  hook's uncommitted-changes check; an in-progress bead tripped it too.
  Now run from an isolated cwd. Passed in CI and failed locally, which is
  how they shipped. (nexus-2v0v7)
• Nine eval graders could not observe what they graded. Every
  not-triggered case was type: llm with criteria beginning "Fail if the
  transcript contains a Skill tool call", but llm's focus defaults to
  last_message and the judge never sees tool calls. They now use
  tool_used with max: 0, the negation primitive the README wrongly
  said did not exist. Proven in the corpus's first real run, where an
  llm grader voted FAIL 3/3 against a max: 0 counter reporting the
  skill was never invoked. (nexus-7zup9)