Fixed simulate mode reporting both removal and install of the same package

[djbclark]

Fixes #24.

In DiffPkgOperations() and ManifestPkgOperations(), inserting an install message into the installed/present map set name_arch to NULL because the map took ownership of it as a key. The MapRemove() call that cancels a previous removal message ran after that, so it looked up a NULL key and matched nothing: a package recorded as removed and then installed was reported as both removed and installed. Moved the cancellation above the code that hands name_arch to the map.

Includes a regression test (tests/unit/simulate_mode_test.c) feeding r,foo,, + i,foo,1.2.3, records to both functions; it fails on master and passes with the fix.

Testing:

• macOS (arm64): test passes with fix, fails on master (rebuilt libcf-agent between runs).
• Ubuntu 24.04 (arm64 container, /bin/sh = dash): same pass/fail discrimination.
• x86_64: this PR's CI run.

AI-assisted, reviewed and verified by me before submitting.

🤖 Generated with Claude Code

[djbclark]

Upstream submission package (human-in-the-loop)

Everything below is ready to paste. Per upstream's restart protocol the submission itself is djbclark's act: 1 bug ticket + 1 PR, smallest possible, with reproduce and before/after output.

Preconditions before submitting:

• The relevant decisions on Consolidated judgment calls from the 2026-08 submissions #23 are answered.
• CI green on head commit 9ee0d1bfc (run 32603359614; the prior identical change already passed run 32602383710).
• Branch state check: git log --oneline -2 fix/simulate-pkg-mapremove shows 9ee0d1bfc on top of upstream master b14a380d0 (rebased and verified 2026-08-22; if upstream master moves substantially before submission, re-verify the cherry-pick).

---

Step 1 — File the Jira ticket (northerntech.atlassian.net, project CFE, type Bug)

Summary:

simulate mode reports both removal and install of the same package

Description:

Recording a package removal and then an installation of the same package makes --simulate=diff / --simulate=manifest report both operations instead of the net install.

In DiffPkgOperations() and ManifestPkgOperations() (cf-agent/simulate_mode.c), the MapRemove() that cancels the earlier removal message runs after MapInsert() has taken ownership of the key and the local name_arch was set to NULL, so it looks up a NULL key and matches nothing. Debug builds instead fail assert(str != NULL) in StringHash() once the map outgrows its array stage.

Reproduce: feed pkgs_ops records r,foo,, then i,foo,1.2.3, (the format RecordPkgOperationInChroot() writes).
Before: output contains both Package 'foo [1.2.3]' would be installed and Package 'foo' would be removed.
After: only the install line.

Found while prototyping a JSON output for simulate mode; the defect is pre-existing on master and independent of that work.

---

Step 2 — Stamp the ticket number into the commit

cd ~/src/cfengine-core
git checkout fix/simulate-pkg-mapremove
git commit --amend -m "$(git log -1 --format=%B | sed 's/Ticket: None/Ticket: CFE-XXXX/')"
git push --force-with-lease origin fix/simulate-pkg-mapremove

(Replace CFE-XXXX. Everything else in the message already follows CONTRIBUTING: past-tense title under 80 chars, full-sentence body, Changelog: Title.)

---

Step 3 — Open the upstream PR

From https://github.com/djbclark/core/tree/fix/simulate-pkg-mapremove → "Contribute" → PR against cfengine:master.

Title:

CFE-XXXX: Fixed simulate mode reporting both removal and install of the same package

Body:

Ticket: https://northerntech.atlassian.net/browse/CFE-XXXX

In DiffPkgOperations() and ManifestPkgOperations(), inserting an install message into the installed/present map set name_arch to NULL because the map took ownership of it as a key. The MapRemove() call that cancels a previous removal message ran after that, so it looked up a NULL key and matched nothing: a package recorded as removed and then installed was reported as both removed and installed. Moved the cancellation above the code that hands name_arch to the map.

The regression test feeds r,foo,, + i,foo,1.2.3, records to both functions: it fails before the fix and passes after, verified on x86_64 Linux (CI), Ubuntu 24.04 arm64, and macOS arm64.

AI-assisted; reviewed and verified by me before submitting, and I take authorship responsibility.

---

Notes for the submitter

• Keep any follow-up replies terse (their explicit earlier feedback).
• Do not volunteer the simulate-json question in this PR — per plan, that probe comes as a short question after this PR has been reviewed.
• The commit intentionally contains the fix and its test together (CONTRIBUTING: tests for the same change share the commit). Squash nothing; it is one commit.
• Evidence links if asked for testing output: fork CI runs above are public; the container/macOS discrimination runs are summarized in simulate mode reports both removal and install of the same package #24 and MapRemove-after-NULL bug (#24) argued at three zoom levels: upstream-terse, full-detail, and auditable-without-C #26.

[djbclark]

Submitted upstream 2026-08-22: ticket CFE-4742, PR cfengine/core#6332 (same commit, 5ddb6d97b = green 9ee0d1bfc + the Ticket: trailer). This fork PR stays open as the CI-evidence anchor until upstream resolves.