fix(resolve): render base branch in summaries and repair BRANCH_INFO phase DAG

[dean0x]

Summary

• Resolution summaries were rendering a literal {base} token instead of the actual base branch name — every resolution-summary.md ever produced by /resolve contained the unresolved placeholder.
• Four phase-dependency annotation defects in resolve.mds repaired so the command's internal DAG is self-consistent and matches the sibling commands (bug-analysis.mds, code-review.mds).

What Was Broken

1. Unrendered template token (user-visible)

The resolution-summary template contained **Branch**: {branch} -> {base}, but Step 0b's extract list never declared base_branch as an output variable. Every generated resolution-summary.md rendered {base} literally. Fixed by extracting base_branch in Step 0b and renaming the template token to {base_branch}.

Sibling-file evidence: bug-analysis.mds:247 carries the byte-identical template line and works correctly because its Step 0b already extracts base_branch. The Git agent's validate-branch op returns **Base**: {base_branch} (src/assets/agents/git.md:177-181) — the value was available and being discarded.

2. Dangling PR_INFO requirement (PF-024 class)

Phases 1b and 9c declared **Requires:** PR_INFO, but PR_INFO has no producer anywhere in resolve.mds or its MDS partials. Renamed to BRANCH_INFO, which Step 0b does produce and which already bundles pr_number (matching bug-analysis.mds).

3. Over-declaration of BRANCH_INFO

Phases 2 and 4 declared BRANCH_INFO but consume no branch value: triage.md declares no branch input, and Phase 4 Code spawns with CREATE_PR: false + PUSH: false, making code.md's only BASE_BRANCH use unreachable. Removed from both.

4. Under-declaration at real consumption sites

Phases 5 and 10 consume BRANCH_INFO (Phase 5 also consumes TARGET_DIR) without declaring it in **Requires:**. Both declarations added, ordered to match code-review.mds:274 and bug-analysis.mds:239.

Net effect

BRANCH_INFO is now produced once (Step 0b) and required at exactly five genuine consumption sites. The DAG is internally consistent. This is a PF-024-class seam fix — the command→agent boundary is an untyped, unchecked call site where annotation rot is the only guardrail.

Verification

• npm run build:mds — 13 commands compiled, 0 errors, 0 warnings.
• 597 tests across 23 files passing (tests/resolve/, tests/ambient.test.ts, tests/build-mds.test.ts, tests/agent-name-guards.test.ts, tests/decisions/, tests/bug-analysis/).
• No test pins any of the changed literals.

Related Issues

None.