feat(mcp): Step 2.5 — input gate + sweep metrics

[2233admin]

Layers on top of PR #6. Implements the three items borrowed from the external agent-memory-runtime governance review (governance-layer.md + decision-table.md):

1. writeAIOutput input gate — reject body<50 chars / single shell command as parent-query / empty query+empty sourceNodes. Placed after schema+enum validation so those errors still surface first.
2. sweep metrics — every sweep response now carries {totalEntries, byPersona, byStatus, byQuarantineState, realBacklinkHitRate}. Signals future threshold-tuning needs: a vault where <10% of entries are cited by human notes is a persona-prompt problem, not a staleness-threshold problem.
3. default-reject rules documented — governance-layer §七 mapped to our vault-level base in docs/ai-output-convention.md.

Skipped from the external review (and why)

• capture/observation layer (append-only log for low-signal persona thought). They need it because their base is Bash-PostToolUse streams; we already filter at the writeAIOutput boundary. Double-layering would be redundant.
• durable_memory + UserPromptSubmit injection gate. Their gate is a Claude Code hook; our recall is a librarian skill reading vault_search. Different interface surface — Step 3 should put relevance/trust logic in the librarian, not a new frontmatter field.

Verification

• npm run build: 0 tsc errors
• ai-output suite: 22/22 pass (6 new tests: 3 gate rejections, 1 gate allow, 2 metrics cases)
• Full mcp-server suite: 110/110 — no regressions

Stats

~180 insertions across 3 files (code + tests + docs). All existing short-body tests updated to use a shared LONG_BODY constant that clears the 50-char gate.

Merge order

PR #6 (feat/step2-governance) first, then this. Both target merges back through feat/ai-output-sediment → v2-staging → main.

[gemini-code-assist]

Code Review

This pull request introduces an 'Input gate' to filter low-signal AI outputs and adds detailed metrics to the sweep process. The input gate enforces quality by rejecting entries with short bodies, shell commands as queries, or missing anchors. The sweep metrics provide insights into persona distribution, status, and backlink hit rates. A performance bottleneck was identified in the metrics calculation where repeated vault scans could cause significant latency in large vaults; a pre-indexed backlink lookup was suggested to optimize this to linear complexity.

[gemini-code-assist]

The current implementation of metrics calculation has a significant performance bottleneck. Calling hasRealBacklink(e.relPath) inside a loop over all entries results in $O(N \times M)$ complexity, where $N$ is the number of AI-Output entries and $M$ is the total number of markdown files in the vault. Since hasRealBacklink performs a full vault scan via walkMd and reads every file, this will lead to severe latency in larger vaults.

I suggest building a backlink index once by walking the vault a single time before the loop. This reduces the complexity to $O(N + M)$. I've also included a minor fix to ensure consistent handling of empty strings in the metrics keys (matching the logic used for status).

Suggested change

	let withRealBacklink = 0;
	for (const e of entries) {
	metrics.byPersona[e.persona] = (metrics.byPersona[e.persona] ?? 0) + 1;
	metrics.byStatus[e.status || "(none)"] = (metrics.byStatus[e.status || "(none)"] ?? 0) + 1;
	const qs = typeof e.fm["quarantine-state"] === "string"
	? (e.fm["quarantine-state"] as string) : "(none)";
	metrics.byQuarantineState[qs] = (metrics.byQuarantineState[qs] ?? 0) + 1;
	if (hasRealBacklink(e.relPath)) withRealBacklink++;
	}
	const backlinked = new Set<string>();
	this.walkMd((relPath, content) => {
	if (aiOutputPaths.has(relPath)) return;
	for (const l of this.parseWikilinks(content)) {
	const link = l.link.split("#")[0];
	if (link) {
	backlinked.add(link);
	if (!link.endsWith(".md")) backlinked.add(link + ".md");
	}
	}
	});
	let withRealBacklink = 0;
	for (const e of entries) {
	metrics.byPersona[e.persona] = (metrics.byPersona[e.persona] ?? 0) + 1;
	metrics.byStatus[e.status || "(none)"] = (metrics.byStatus[e.status || "(none)"] ?? 0) + 1;
	const qs = (typeof e.fm["quarantine-state"] === "string" && e.fm["quarantine-state"]) || "(none)";
	metrics.byQuarantineState[qs] = (metrics.byQuarantineState[qs] ?? 0) + 1;
	if (backlinked.has(e.relPath) || backlinked.has(basename(e.relPath, ".md"))) withRealBacklink++;
	}

[2233admin]

Collapsed into PR #4 for unified v2 merge (squash into main). feat/step2-6-tag-migration fast-forwarded into v2-staging; all commits in this PR are now part of #4's diff.