2.2 KiB
#1500 prerequisite correction, pre-review
Jason approved the narrow corrective slice by answering yes to the request for D1-D10 repairs, adversarial tests, independent review and selective commit/push. This does not authorize materialization or live operations. Baseline b3fa221060.
Implemented only in packages/mosaic:
- Fail-closed required roots/directories and empty registries; no partial results.
- Linux descriptor-anchored traversal, lstat and O_NOFOLLOW checks for metadata and directories, inode/device rechecks and private owner/modes. Unsupported platforms refuse. No external dependencies.
- Numeric version 1, all-harness strict nested metadata, URL userinfo/protocol refusal and explicit HTTP opt-in.
- Default enrollment, provider credential-type and reference checks.
- Bounded JSON reads and fixed diagnostics, no parser-content excerpts or input-key/value reflection.
- Numeric fixture migration, private temporary fixture copies, D1-D10 regressions, null/malformed data and unreadable credential-sibling controls. No real credential files read.
Verification: 43 package tests passed; combined package/launcher command passed 48/48. Syntax checks and diff whitespace check passed. Logs /tmp/i1500-safety.log and /tmp/i1500-combined.log. An earlier package run was 14/15 because the missing-provider control received empty-registry first; reordered reference validation before final emptiness check, retained original failure at /tmp/i1500-initial.log, rerun green. No missing registry is accepted in either version of that correction.
Limits: Linux/procfs is explicit in package README; no claim of transactional multi-record snapshot consistency under concurrent writers. Credential sibling is never opened/stat'ed. No root pin/dependency changes or launch-service edits. Docker/task/release suites have not been rerun for this package-only correction; tests with image/state side effects are not implied authorized by their names. Prior full-task failure is not declared fixed. No code committed or pushed before independent review.
Next: Filbert independently reviews exact package overlay and reproduces adversarial controls in a clean baseline copy. No dependent materialization work until approval and correction publication.