Remove the debt artifacts and put the Rust suite under an automatic gate

[niveousdragon]

The project is in release stage, where technical-debt artifacts do not belong in
the tree. This reverts the part of #283 that recorded debt, and replaces the
mechanism it stood in for: instead of capping three debt classes at their
current size, the Rust test suite now runs under a gate that catches whatever
grows.

Removing the debt artifacts

#283 added 1725 lines. 1696 of them recorded debt rather than removed it:

• four audit ledgers and a remediation plan in docs/;
• python/tests/test_structural_debt_ceilings.py, whose ceilings of 105
  unreferenced public functions, 185 silent WCOJ declines and 107 environment
  names assert that this much debt is an acceptable release state;
• a paragraph in ENGINEERING.md saying the same inside the binding contract.

All of it is gone. Findings are tracked outside the release repository, one
issue per finding, and each is closed by the change that fixes it — the counts
go to zero, and the tests that hold them there assert zero rather than a
ceiling.

The wording about the two kinds of WCOJ decline is reverted with the rest. It
states an observability requirement the dispatcher does not meet today, so it
belongs in the same change as the counter that satisfies it, not ahead of it.

.cargo/config.toml from the same pull request is kept: it repairs a real
defect — cargo unusable outside Linux — rather than recording one.

Putting the Rust suite under a gate

cargo test --workspace --all-targets --release ran under workflow_dispatch
alone, so it ran when somebody remembered to press the button. That is the
mechanism by which a broken test once sat on main until an outside run
happened to find it.

• Nightly. rust-tests now runs on a schedule (0 2 * * *) against main,
  in addition to workflow_dispatch.
• Every pull request. A new cuda-slice job runs
  cargo test --all-targets --release -p xlog-cuda -p xlog-gpu on the CUDA
  runner — the two crates whose entire purpose is the device, including the
  certification suite in xlog-gpu/tests. The whole workspace is too long for a
  pull request; this part is not.
• No path filter on cuda-slice, deliberately. A filter deciding which pull
  requests get GPU coverage leaves whatever it misses unchecked, which is the
  same class of mechanism as the button above.
• Failures reach a person. report-scheduled-failure opens an issue, or
  comments on the open one when a previous night already failed. A nightly run
  whose failure reaches nobody is a cron job, not a gate.

Two defects surfaced while wiring the schedule and are fixed here. The change
classifier rejected any event it did not recognise, so a scheduled run would
have failed in its first job before reaching the tests. And the concurrency
key grouped by ref alone, so a push to main would have cancelled a nightly run
already an hour in.

Verification

• python/tests/test_ci_workflow_contract.py gains a test asserting the whole
  shape: the cron entry, schedule in the rust-tests condition, cuda-slice
  on the CUDA runner for same-repository pull requests, the absence of a
  path filter on it, the issue-reporting job's issues: write permission, and
  the event name in the concurrency group.
• Both negative checks pass: removing schedule from the rust-tests
  condition fails the test, and adding needs: cuda-changes to cuda-slice
  fails it. Tree restored after each.
• evaluate_python_wheel_gate simulated for schedule: passes on GPU success,
  fails on GPU failure.
• Contract suite: 16 passed, 1 failed. The failure is
  test_cuda_change_classifier_cli_consumes_null_delimited_git_paths, which
  fails on main too and only off Linux (CRLF from print()); it is a separate
  finding with its own fix, not introduced here.
• Against the tree as it stood before Publish the debt ledgers, cap three debt classes, and unbreak cargo off Linux #283, exactly one file differs:
  .cargo/config.toml. No dangling references to the deleted files remain.

Not measured: how long cuda-slice takes. No GPU on the machine this was
written on. The 60-minute timeout is generous on purpose; this pull request is
its first real run.

Follow-up that is a ruleset change, not a commit

protect-main requires nine checks and none of them is GPU-facing. cuda-slice
and python-wheel-gpu should be added — the latter already runs on pull
requests without blocking — along with changelog-ownership.

One subtlety worth recording: python-wheel-gpu reports "skipped" when the
classifier says no CUDA inputs changed, and GitHub counts a skip as a pass. That
makes scripts/cuda_ci.py load-bearing for a required check. cuda-slice has
no such dependency, by design.