Report when the gold check had nothing to check against

[ATMAECHO]

The problem

Task.validate returns True unconditionally, so a taskset that does not override it
passes the gold check for every item without anything having been checked.

Of the five in-tree tasksets defining a Task subclass, only lean overrides validate.
harbor, openenv, textarena and nemo_gym inherit the base implementation. For those,
validate --only-gold reports valid_rate: 1.0 — which reads as a passing check, but
measures nothing. The failure is silent and flattering: absence of a check is reported as
agreement.

The change

34 lines in verifiers/v1/cli/validate.py:

• gold_implemented(task) — whether the taskset actually wrote a gold check
  (type(task).validate is not Task.validate, so inheriting from an overriding class
  still counts).
• _run_gold marks each row checked: true|false.
• summarize reports gold_unchecked in gold and all modes (_gold_row reads the
  nested gold row in all mode).
• run_validate warns once, naming the classes that don't override validate.

Validation behaviour is unchanged. This only separates "checked and valid" from
"never checked".

A stronger alternative, if you'd prefer it

Add an unchecked value to reason so it fails loudly rather than being reported in the
summary. I kept this additive because that would touch FINAL_REASONS, _is_final and
resume semantics — happy to switch if you'd rather have the louder version.

Verification

Per AGENTS.md ("to check your own work, write a temporary script instead of committing new
tests") I used a temporary script, not a committed test. It checks:

• the predicate on synthetic cases: inherits Task.validate, overrides it, and inherits a
  class that overrides it;
• all five in-tree taskset classes, imported live — 4 unchecked, lean checked;
• summarize in gold, all and setup modes, including a row with no checked field;
• a negative control — the old always-true predicate — which fires, so the checks aren't
  passing vacuously.

All pass. uv run ruff check and uv run ruff format --check are clean.

One caveat I'd rather state than hide: I could not run uv run pytest tests/ on this
machine. verifiers/v1/runtimes/limiters.py imports fcntl, which is Unix-only, so the
package doesn't import on Windows. I stashed the change, collected, un-stashed, and
collected again — the failure set is byte-identical (38 both ways, same 6 collection
errors), so nothing here is caused by this change. Still worth a CI run on Linux.

Note

Report unchecked gold validations when a task does not override Task.validate

• Adds a gold_unchecked count to the summary in gold and all modes, tracking results that passed without a real gold check being implemented.
• Each result row from _run_gold now includes a checked boolean derived from whether the task's class overrides Task.validate.
• Before validation runs, run_validate logs a warning listing any task classes whose gold checks will always pass due to missing validate overrides.

^{Macroscope summarized 4a74e93.}

[xeophon]

superseded by #2266