Kernel: a reviewing Run that leaves a rubric criterion uncovered has not reported
Closes #16.
A reviewing Run that leaves a criterion of the rubric uncovered has not reported. It doesn't count as the model review a merge needs, and the existing attempts brake bounds the retries. The rulings are on #16, and decision 0029 records them.
What changes
kernel/coverage.py(new, protected): one pure function,covered, decides covered, uncovered (with the criterion ids) or unreadable. It judges the rubric's criteria against the Run's coverage list and Findings, and the changed paths, including both names of a rename.coverage/1: a schema of its own for the coverage list, with an example and a counter-example. #19 is left waiting on its trials.- No new fact: coverage lives in the Result, which the Kernel never reads. The run record carries the outcome,
conclusion: failurewithoutcome: item. Only a successful review Run counts as a report. Failed ones are re-admitted, and after 3 attempts the item is held. - New reason,
run_unreadable: a "success" Run that also carries a fault, from a required Lens's Role, defers to the Maintainer. A queued item leaves the queue. **in the one path syntax: it now means any number of names, in bothkernel/derive.pyandtools/hooks/protected.py, and the agreement test covers it. The hooks' check for a directory holding a protected path doesn't handle**yet, soload()refuses**in any protected, reported or Lens-input entry until it does.- Simulator: uncovered reviews occur on one seed in six.
Found while building, filed separately: #158. "Kernel: spent attempts hold the item even while a refused Run is queued". With attempts spent, a queued Run that the brakes would refuse hides the Hold, so the Factory moves the item from held to verifying, a move the Lifecycle doesn't declare. No Effect and no model spend follow. This predates this story. The simulator's narrower spread, UNCOVERED_SEEDS, carries a comment naming #158, which restores it.
Verified:
make verifypasses (540 tests) on a plaingit archiveexport withPATH=/usr/bin:/bin. The record was renumbered there only, for the gap; this PR is red in CI only on that gap until 0027 and 0028 merge.- Each new mutant was checked by hand against its own test, and is killed by an assertion. No local mutation run; CI's shards run the set.
check_decisionsis clean, and the simulator passes at 300 seeds.
Reviewed by: a fresh-context Claude review, a security pass, and AGY (Gemini 4 Argon) on c66e1ee. Nothing found lets an uncovered review reach an approval. Their findings are fixed in ed9b014:
- the criterion-id checks, the required-Role guard and each clause of
contradictshad no test or mutant; - the dedupe and readability checks were untested;
**and the directory-holds check.
Left, informational:
coverage.pyis protected but imports the unprotectedderive.matches.kernel/is a Lens input at PR time, so a change to it needs every model Lens's review (0020).- A Finding that cites a criterion not in the rubric still gives
covered. - Until
applycallscovered, coverage is enforced by whoever writes the run record. That is the harness row in 0029.
Order: merges after #155 (decision 0028).
š¤ Generated with Claude Code