Date: 2026-08-28
Repository: MSKazemi/idkmesh
Continue
Repository context: https://github.com/MSKazemi/idkmesh
By this point the repository had converged substantially:
main for frozen node PR #91;The frozen node head remains:
d638a2f78e4a89353b98e91052233e365f56f90a
This continuation did not move that SHA.
PR #110 automated the positive controlled-host path but intentionally left the five issue #37 negative runtime cases manual.
That preserved safety, but it also left a high-friction community task: a contributor would need to construct temporary WorkUnit variants correctly, distinguish expected worker failure from harness failure, retain evidence, and avoid leaving test state behind.
The next useful repository-internal improvement was therefore:
make the negative runtime matrix reproducible without reducing it to mocked CI or changing the candidate under test.
run-negatives commandscripts/pr91_acceptance.py now adds a controlled-host command:
run-negatives
The command:
/tmp rather than committing destructive fixtures;A temporary WorkUnit asks the real node to produce a tracked change outside the canonical allowed path.
The harness requires the worker to fail closed while still retaining the changed path and path-policy evidence.
The purpose is to prove that a task cannot turn an out-of-scope tracked change into a successful candidate merely because the task process itself exits normally.
The source revision ignores the results/ subtree.
A temporary WorkUnit allows that subtree for path/write scope, then emits an ignored untracked file. This isolates the artifact-completeness rule from ordinary path-policy rejection.
The harness requires:
This is the runtime counterpart to the node’s fail-closed rule that .gitignore cannot make candidate output disappear from evidence.
A temporary task attempts to alter the task-visible Git pointer while also making a normal allowed tracked change.
The expected invariant is not that the task-visible pointer is magically immutable. The invariant is that the candidate cannot control the host-owned Git metadata used to measure the candidate.
The harness therefore requires:
This tests the separation introduced by PR #91 between task-writable workspace state and host-owned read-only Git metadata.
A temporary WorkUnit uses a deliberately small candidate-patch byte bound and creates a larger otherwise-allowed tracked change.
The harness requires:
The point is that an incomplete candidate artifact may be retained diagnostically but cannot become a successful candidate.
The node allows two small image tags in this MVP. The acceptance host is instructed to preload only the positive python:3.12-alpine candidate image.
For safe automation, the harness refuses to replace any pre-existing alternate allowlisted Docker tag.
On a clean controlled host it then checks two failure modes:
The temporary tag is removed in a guaranteed cleanup path and absence is rechecked. If the alternate tag existed before the run, the harness refuses the test rather than mutating that host state.
A–D are expected to return a normal worker-attempt failure class with a retained ResultManifest/evidence bundle.
E is expected to be refused before normal candidate execution because immutable image evidence is missing or mismatched.
The aggregate report records:
exact frozen SHA
host / Docker / image preflight
case A..E command observations
case A..D retained evidence validation
case E image-identity observations + cleanup state
per-case passed/errors
aggregate passed/failed_cases
automatic_merge = false
integration_decision = null
A report is evidence for issue #37 human review. It is not itself authority to close the issue or merge PR #91.
GitHub-hosted CI still does not run Docker for this acceptance gate.
The existing harness check continues to compile the script and run its Docker-free deterministic self-test. The self-test now also checks the negative WorkUnit mutations and repository-digest mismatch logic without pretending those synthetic checks are runtime acceptance.
current main acceptance harness
-> separate exact PR #91 checkout
-> positive controlled Docker run
-> A-E controlled Docker negative matrix
-> issue #37 human evidence review
-> real bundle replay through patch verifier v0.1.1
-> VerificationResult + Evidence Report/replay
-> human integration decision
-> first 5-10 benchmark tasks
No self-merge or autonomous integration is introduced by this continuation.