chore(agent-memory): record completion-gate receipt shape and collect_pr_context defects (#445 follow-up)

[drmoisan]

Summary

Agent-memory only. Follow-up to PR #587 (epic child #445), which is already merged into this integration branch. No production, test, or policy file is touched.

Records two durable findings from the #445 run, both correcting earlier memory that was pessimistic or incomplete.

The completion-gate receipt shape is no longer undiscoverable

Get-CheckpointAcknowledgedName in .claude/lib/orchestrator-state/OrchestratorStateRoutingContract.psm1 reads a hard-coded evidence member, so a skill or MCP receipt counts only when all three of these hold: its name is a non-blank string, its flag is boolean true (truthy is not enough), and it carries a non-blank evidence string.

Earlier runs guessed at alias key names (skills_used, mcp_receipts, mcp_tool_calls) and never supplied evidence, which is why the gate kept reporting the same receipts missing. That is why the prior memory concluded the shape was undiscoverable from outside the tool. It is discoverable — by reading the module rather than probing the validator.

The bug-route tool-name swap, and the divergence it creates

Get-ResolvedRequiredMcpTool swaps new_potential_entry for new_potential_bug_entry when the promotion type is bug. The MCP TypeScript surface does not apply that swap. Because required_mcp_tools is compared by exact list equality, no single list satisfies both validators on a bug route.

The hook is authoritative, so the guidance is to conform to it and expect two residual MCP errors that are divergence artifacts rather than missing work.

Together these supersede the earlier conclusion that --require-complete is simply unsatisfiable for a resumed-at-execution epic child. It is satisfiable at the real gate: .claude/hooks/validate-orchestrator-output.ps1 returns EXIT_CODE 0 on the #445 checkpoint.

collect_pr_context has two defects beyond misplacement

The existing entry recorded that the tool writes into the primary checkout. Two further defects are added, because they make the artifact unusable rather than merely misplaced: it reports GitHub CLI unavailable when gh is authenticated in the same worktree, and the primary checkout it writes into sits on a different branch, so the diff it computes is not the branch under review.

The remedy is therefore to author the PR body from real git data, not to copy the artifact into the worktree as the entry previously advised.

Verification

No GitHub Actions workflow runs on this pull request: .github/workflows/ci.yml triggers pull_request only on [main, development] and this PR targets the epic integration branch. The absence of checks is expected.

No toolchain stage is re-run, and none is required: the diff is three Markdown files under .claude/agent-memory/orchestrator/, which no build, analyzer, or test consumes. The C# toolchain result established for PR #587 against commit f912640b is unaffected.

Scope

• Only .claude/agent-memory/orchestrator/** is modified — the one subtree the epic brief carves out of the .claude/** prohibition.
• No rule, skill, hook, or config file is touched.
• No production or test source is touched.

🤖 Generated with Claude Code