fix(gbrain): cap fallback enrichment graph size (BLO-21635)

[kkroo]

Supersedes #1070, whose mixed history is rejected by the App-attributed-commit policy.

Reapplies the same final two-file gbrain fallback cap onto current master as one kkroo-authored commit. pnpm-lock.yaml is unchanged.

Validation:

• Isolated pnpm install --frozen-lockfile on the original fix(gbrain): cap fallback enrichment graph size (BLO-21635) #1070 head
• Focused gbrain recall tests: 21 passed
• Gbrain typecheck passed
• git diff --check passed

The full local gbrain suite had one pre-existing recall-pack decompression test timeout; the changed recall.test.ts suite passed. GitHub CI remains the merge signal.

[allyblockcast]

🔗 Paperclip issue: BLO-21635

[allyblockcast]

🔗 Paperclip issue: BLO-21635

[allyblockcast]

Hey @kkroo! Before this PR can be reviewed, a few things need attention:

Missing or incomplete:

• Missing section: ## Thinking Path
• Missing section: ## What Changed
• Missing section: ## Verification
• Missing section: ## Risks
• Missing section: ## Model Used
• Add the dedup-search checkbox to your PR description and check it once you have searched the GitHub PR list for similar PRs. See the PR template at .github/PULL_REQUEST_TEMPLATE.md and CONTRIBUTING.md → "Before You Start: Search First".

Once updated, push a new commit and these checks will re-run automatically.

— commitperclip

[allyblockcast]

Hey @kkroo! Before this PR can be reviewed, a few things need attention:

Missing or incomplete:

• Missing section: ## Thinking Path
• Missing section: ## What Changed
• Missing section: ## Verification
• Missing section: ## Risks
• Missing section: ## Model Used
• Add the dedup-search checkbox to your PR description and check it once you have searched the GitHub PR list for similar PRs. See the PR template at .github/PULL_REQUEST_TEMPLATE.md and CONTRIBUTING.md → "Before You Start: Search First".

Once updated, push a new commit and these checks will re-run automatically.

— commitperclip

[kkroo]

Closing as superseded: this exact change is already on master.

Commit a73d8500 — fix(gbrain): cap agent/project fallback enrichment graph size (BLO-21635), landed 2026-08-06, i.e. four days before this PR was opened — carries the identical cap in packages/plugins/paperclip-plugin-gbrain/src/recall.ts:

// BLO-21635: the fallback candidates are agent/project hub nodes, which can be
// linked to thousands of unrelated fact-* nodes (observed up to ~23k, ~5.8MB
// decoded JSON, on a single "ok" recall). The issue traversal is scoped and
// safe to attach whole; the hub fallback is not, so cap only the fallback's
// contribution before merging rather than letting it dominate the cached
// graph. This bounds payload size regardless of how big the server-side hub
// neighborhood is.
export const MAX_ENRICHMENT_NODES = 50;
export const MAX_ENRICHMENT_EDGES = 100;
export const MAX_ENRICHMENT_SERIALIZED_BYTES = 64 * 1024;

function mergeGraphs(primary: unknown | null, fallback: unknown): unknown {
  const cappedFallback = capEnrichmentGraph(fallback);
  ...

…along with capEnrichmentGraph, capNodesArray, nodeRank, nodeSlug, edgeWithinSlugs, capArrayBySerializedBytes, capGraphRecord, and serializedBytes — and all four cap tests in src/__tests__/recall.test.ts (caps an oversized agent-hub fallback (array shape)…, …(nodes/edges shape)…, normalizes and caps an oversized edges-only fallback graph, drops auxiliary fallback collections and caps dense kept-node edges).

Verification against origin/master @ 3c59d88d:

$ git diff origin/master codex/blo-21635-gbrain-enrichment-cap-successor \
    -- packages/plugins/paperclip-plugin-gbrain/src/__tests__/recall.test.ts
(no output — test file is byte-identical)

$ git diff origin/master codex/blo-21635-gbrain-enrichment-cap-successor \
    -- packages/plugins/paperclip-plugin-gbrain/src/recall.ts | grep '^+' | grep -v '^+++'
+    return { ok: false, issuePageSlug: slug, graph: null, reason: `traverse_graph failed: ${msg}` };
+    const status = input.result.issuePageSlug ? "error" : "skipped";

The only two lines this branch would contribute are the pre-BLO-23403 forms of those statements — i.e. master is strictly ahead here, having since added the no-oauth-client reason-kind handling on top of the cap. This branch has nothing left to add.

Note: this closes the PR only. It was not verified whether BLO-21635 is fully resolved in production — only that the code and tests this branch proposed are already merged.