Push-gate execution evidence

ci_gate_replica.py is the single producer for workflow-derived heavy checks in the pre-push gate. It parses the checked-in workflow files and builds one bounded plan. Identical selections in that plan execute once per invocation; later consumers may reuse only a verified result from that same invocation (or an explicitly declared completed prior phase).

This is an execution optimization, not a coverage waiver. A missing, partial, stale, failed, tampered, unsigned, or differently configured record always falls through to the normal command.

Evidence identity

Each selection is keyed by a canonical digest of:

• the exact command argument vector and command kind;
• normalized packages, features, targets, and other flags extracted from the vector;
• the effective Cargo environment, including target directory and bounded parallelism.

The invocation source identity includes the committed revision and tree, working-tree and index binary diffs, bounded untracked-file content, the Cargo.lock digest, and pinned toolchain/configuration plus rustc/cargo version output. A successful result is additionally bound to its selection payload and exit code. Evidence is stored in the worktree's private Git directory, with restrictive permissions and a per-invocation HMAC key. It is never a release artifact and contains environment digests rather than raw environment values.

The pre-commit coordinator's PID and Linux process-start token identify the invocation across the per-hook shell wrappers; bounded pre-commit stage/ref fields further distinguish separate pushes under a long-lived parent. Records older than 24 hours, records from another process, or records whose source/configuration identity changes are not admissible.

Admissibility predicate

A record can be consumed only when all of the following hold:

1. the evidence schema, invocation identity, source fingerprint, file modes, content digest, and HMAC verify;
2. the invocation status is complete;
3. the selection payload is present exactly in the plan;
4. the result is success, has exit code 0, and has the expected result digest.

The sole non-exact relation is the versioned cargo-clippy-full proof. The advisory cargo clippy --workspace --all-features --all-targets -- -D warnings result may cover the shipped full/all-targets command only when the requested command, provider payload, and effective environment match the declared proof. No other selection is inferred to be a subset, and mandatory coverage remains mandatory.

Failure and restart behavior

The producer marks a resumed cache running before executing any new plan item. It records failed results as failed and leaves them non-consumable. An interrupted or otherwise non-finalized invocation remains partial and is executed normally on the next attempt. A corrupt marker, key, or evidence file creates no pass path: the cache is ignored and the command runs normally. The evidence plan also has a hard selection-count bound; an unexpectedly expanded workflow disables reuse rather than allocating without limit.

The constrained two-core gate is intentionally a different environment and coverage selection. Its CPU-affinity/runtime proof is not silently substituted for the workflow replica's ordinary Cargo selection.

Root validation

The landing agent should inspect the complete source-level plan and run the focused admissibility tests before the repository's normal pre-push/CI gates. The focused selector is:

pytest -q tests/test_push_gate_evidence.py

The source-only worker did not execute this command or any Cargo/build/hook command.