Hard-reset integration branch after merge to prevent SHA divergence

[sebastientaggart]

Replaces git pull on the integration/production branches with git fetch + git reset --hard origin/<branch> so the local branch becomes a perfect mirror of the remote after every CodeCannon-driven merge.

The previous git pull after gh pr merge --merge was creating a second, parallel merge commit locally for the same logical PR — once any divergence existed, every iteration compounded it (the symptom in the linked issue: 466 vs 465 different commits, near-empty content diff). Hard reset is safe because the workflow forbids direct commits to the integration/production branches, so there is no local work to lose (and git reflog keeps a 90-day safety net regardless).

Changes:

• Makefile.agents.mk merge: target — git pull → git fetch + git reset --hard origin/$(INTEGRATION_BRANCH).
• skills/start.md Step 4 (Case A and Case B, both BRANCH_DEV and BRANCH_PROD variants) — same pattern.
• skills/deploy.md Step 1 sync — bare git pull → git fetch && git reset --hard @{u} (uses upstream so it works for whichever branch the project's deploy mode requires).
• Rendered skill outputs synced.

Closes #156

[sebastientaggart]

Review Summary

Verdict: APPROVE

The change faithfully implements the fix described in #156: replaces git pull with git fetch + git reset --hard origin/<branch> in the four sites that sync the integration/production branches (Makefile.agents.mk merge: target, skills/start.md Case A and Case B Step 4, skills/deploy.md Step 1). All four rendered adapter outputs (.claude/, .cursor/, .agents/, .gemini/) match the templates and updated content hashes. The acceptance criteria from #156 are satisfied by the diff.

Findings

• [NOTE] skills/deploy.md uses git reset --hard @{u} while skills/start.md and Makefile.agents.mk name the branch explicitly as origin/<branch>. Both forms are correct given the preceding git checkout / required-branch guard, but using @{u} introduces a small inconsistency and depends on the local branch having an upstream configured pointing at origin/<deploy-branch>. Naming the branch explicitly (matching the other two sites) would be marginally more defensive and stylistically consistent.

• [NOTE] git reset --hard silently discards any uncommitted working-tree changes on the integration/production branch, whereas the previous git pull would have aborted loudly on a dirty tree. This is safe under the documented CodeCannon workflow (integration/prod branches are never edited locally — all changes flow through feature branches + PRs) and Local integration branch (dev) accumulates SHA-divergent history vs origin/dev over time — make merge uses git pull where it needs git reset --hard #156 already calls this out, with git reflog providing 90-day recovery. Worth being aware of as a behavior change for any user who had been hand-editing those branches against convention.

[sebastientaggart]

Review Summary (re-review after 34c1f16)

Verdict: APPROVE

The new commit cleanly addresses both prior NOTEs and adds the local-only-commits warning specified in #156's recovery rationale.

Verified against the diff:

• Dirty-tree guard (git diff --quiet && git diff --cached --quiet || abort) is present at every git reset --hard site: Makefile.agents.mk merge:, skills/start.md Step 4 Case A and Case B (both BRANCH_DEV and BRANCH_PROD variants), and skills/deploy.md Step 1 (all three conditional blocks).
• Local-only-commits warning in Makefile.agents.mk counts git rev-list --count origin/$(INTEGRATION_BRANCH)..HEAD after git fetch and prints git log --oneline origin/$(INTEGRATION_BRANCH)..HEAD before the reset. Non-aborting, as intended for the Local integration branch (dev) accumulates SHA-divergent history vs origin/dev over time — make merge uses git pull where it needs git reset --hard #156 recovery path. Output reaches the user's terminal via the make recipe.
• skills/deploy.md three-mode conditional now covers all three branching modes: {{#if BRANCH_TEST}} → reset to origin/{{BRANCH_TEST}}, {{#if !BRANCH_TEST}}{{#if BRANCH_DEV}} → reset to origin/{{BRANCH_DEV}}, {{#if !BRANCH_TEST}}{{#if !BRANCH_DEV}} → reset to origin/{{BRANCH_PROD}}. Matches the trunk/two-branch/three-branch documentation block at the top of Step 1.
• The earlier @{u} form is gone — every site now names the branch explicitly as origin/<branch>, resolving the stylistic NOTE from the prior review.

Findings

• [NOTE] The dirty-tree guard git diff --quiet && git diff --cached --quiet deliberately ignores untracked files. This is the right call — git reset --hard does not touch untracked files, so guarding against them would produce false aborts (e.g. editor swap files, build artifacts, .DS_Store) without protecting any user work.

• [NOTE] The guard can produce a benign false abort if the working tree has uncommitted submodule pointer changes, file-mode flips (e.g. on filesystems where core.fileMode differs), or CRLF/autocrlf renormalization noise that leaves git diff non-empty without user-visible edits. The failure mode is loud (clear error message asking the user to resolve) rather than silent data loss, so this is the correct trade-off; flagging only for awareness.

• [NOTE] The local-only-commits warning in Makefile.agents.mk is informational rather than interactive — by the time the user reads "discarding N commit(s)", the reset has already executed (chained with &&). Recovery via git reflog within 90 days remains available, matching the recovery story documented in Local integration branch (dev) accumulates SHA-divergent history vs origin/dev over time — make merge uses git pull where it needs git reset --hard #156. Acceptable, and prompting for confirmation inside a make recipe would be awkward, but worth noting that the message is post-hoc.