fix: close the verify-gate bypass, reach the last two gates from MCP, correct the docs

[bhyi4]

An audit of every doc claim against the code. Three things mattered; the rest was hygiene.

1. The verify gate could be bypassed by deleting the thing it keys on

verify-gate matched - [x] … verify: \CMD``. An item ticked after its verify clause was removed
matched nothing, so the gate skipped it and the box survived unverified:

- [x] item with verify deleted   →  exit 0   (passed, not reverted)
- [x] item. verify: `false`      →  exit 1   (caught, reverted to [ ])

ralph's eligibility check only looked at unchecked items, and only once, before the loop —
so nothing re-checked it mid-run. The documented guarantee ("a checkbox survives only if its verify
command actually passes when the harness runs it") did not hold on that path.

Fix: --require-verify treats a checked item with no verify clause as a failure and reverts it.
ralph passes the flag and refuses a TODO where any item — checked or not — lacks the clause, which
is what makes the flag safe without exemption bookkeeping. The default stays lenient so a mixed
TODO (loop items + manual items) can still be scanned standalone.

2. The two strongest gates had no MCP surface

arc-prereg (link a seal so arc-close injects the kill-condition verbatim) and verify-gate were
shell-only. An agent driving Yeoul through MCP — the setup the README prescribes — could run neither,
while docs/BOOTSTRAP_PROMPT.md told it the verify gate existed "as a tool, not just instructions".

Added arc_prereg and verify_gate. Tool surface 10 → 12, CI assertion updated.

3. Docs

• docs/METHODOLOGY.md referenced bin/yeoul-graduate; the command is bin/graduate.
• The verify-gate claims in README, README_KO, METHODOLOGY and the bootstrap prompt now say what is
  enforced and what is not: the harness owns the re-run and the presence of the clause, but a verify
  command that cannot fail is still the author's problem.
• Backend A's obligation to run the gate itself is explicit instead of a parenthetical.
• CHANGELOG records this batch plus the two previously merged fixes.

Tests

5 new assertions covering the bypass, the compat path, and no-false-positives. Verified they fail
against the previous scripts — 3 of 5 fail (numbered above); the other 2 are the compat and
no-false-positive controls, which pass on both by design. pre-publish-check.sh and examples/demo.sh
still pass.

Note on that check: its over-claim scan looks for marketing superlatives, so it reported clean
throughout — it cannot catch a claim that contradicts the code. That is why this audit was manual.

🤖 Generated with Claude Code