Guard make targets with repo-root cd and existence check

[sebastientaggart]

The submit-for-review skill runs make targets (CHECK_CMD, MERGE_CMD) in whatever directory the agent happens to be in. If the agent drifted into a subdirectory, the command fails and the agent improvises a partial fallback — silently skipping checks or inventing nonexistent targets like make merge.

Two fixes:

1. cd "$(git rev-parse --show-toplevel)" before running any make command, so it always resolves against the root Makefile.
2. make -n <target> existence guard before execution — if the target doesn't exist, the skill halts with a clear error message instead of letting the agent improvise.

Issue #129

[sebastientaggart]

Review Summary

Verdict: APPROVE

Findings

• [NOTE] The instruction "Extract the target name from make check (e.g. make check → check)" relies on the AI agent to parse the target from a hardcoded string at runtime. This works for the current make check / make merge commands but would break if CHECK_CMD or MERGE_CMD were ever set to something other than a simple make <target> invocation (e.g. make -C subdir check). Acceptable for now since .codecannon.yaml only supports simple make targets, but worth noting for future extensibility.

• [NOTE] The 2>/dev/null on make -n silences stderr, which means if make itself is not installed or the Makefile has a syntax error, the agent will report "target does not exist" rather than the actual underlying problem. The error message could be slightly misleading in that edge case, but the fail-closed behavior is correct.

No blocking issues found. The changes are consistent across all five adapter files, the source template uses {{CHECK_CMD}} / {{MERGE_CMD}} placeholders correctly, and the rendered adapters all resolve to make check / make merge as expected. The cd to repo root and make -n existence guard both address the problem described in the PR body (issue #129) and follow the project's fail-closed philosophy.