2026-06-15 — Per-source `gather()`: fix bulk-corpus retrieval noise drowning curated pages

Status: Implemented (slaude-side wrapper), tested. Upstream promotion to gbrain is a separate follow-up (see "Promotion" below).

Symptom#

"Same question, different answer." In a warm-context thread, a directory lookup ("who is the lead of team X") resolved to the right person. In a cold thread the same question → "the brain does not contain that data." Reproduced cold-vs-warm cleanly.

Diagnosis (traced from session jsonl + live brain)#

Two layers, one root.

1. Warm context masks the bug. A long thread loads the curated directory page early (an earlier lookup), and it stays in the conversation window — so later role questions are answered from context with zero KB calls (verified: the answer fired only a reply, no kb_*). A cold thread starts empty every time and is forced through retrieval.

2. The cold retrieval is poisoned by a bulk auto-generated corpus. That source holds thousands of near-empty # model / no documented columns stubs. On any common domain-word query they score high base-vector and fill the candidate gather by sheer volume — a live kb_think gathered 25 pages, all 25 bulk stubs; the curated page never entered the pool, so the reranker (which correctly tanks the stubs to ~0.06) never got to sort it up. Proof it's volume, not scope: in the same thread, a distinctive-token query (a rare name no bulk model is named after) hit the curated page at rank 1 instantly.

So: distinctive-token queries work anywhere; common domain-word queries fail anywhere (even warm, if context were empty). Warm threads just rarely fire a cold domain-word query.

Root cause#

gbrain pools all allowed sources into ONE ranked candidate set (the search op and think's internal gather both do). A high-volume source monopolises the gather before the reranker runs. Per-source representation is not guaranteed.

Fix — slaude-side per-source gather()#

src/knowledge/gather.ts: fan out one search per allowed source (single-source scope each), take top-K from each, merge + re-rank the union by gbrain's own rerank_score. Every source — including small curated ones — is guaranteed K candidate slots; no source can monopolise by volume. The reranker then sorts the stubs down and the curated page up.

• kb_search → routes through gather().
• kb_think's zero-citation cross-check (#38) → uses gather() (per-source) instead of a single pooled search, so the uncited present page actually surfaces in search_fallback rather than more bulk-corpus junk.
• gbrain stays stock (github:garrytan/gbrain#03ffc6e) — we only orchestrate its search op. No fork.
• Total-failure posture: a single source failing drops to []; if every source fails the brain is down, so gather rethrows → handler maps to an error (never papers over a dead brain with empty results — cf. [2026-06-14-brain-memoize-failure.md]).

Known limitation (follow-up)#

kb_think's primary gather is inside gbrain's runThink — the slaude wrapper can't reach it, so the synthesis still gathers pooled. Today the per-source cross-check rescues the answer via search_fallback; the full fix is to re-point kb_think's candidate step at the shared gather() (compose gbrain's renderPagesBlock + prompt builders + the SDK synth client). Deferred: higher risk (faithful think-prompt reproduction), to be proven separately.

Promotion#

This slaude wrapper is the proving ground. Once validated in prod, promote per-source ranking upstream to gbrain (gather.ts / the search op) so it heals back into the dependency and both entry points gain it natively.

Tests#

tests/gather.test.ts — deterministic injected-call units (fan-out, flood-survival, dedup-by-slug, finalLimit+sort, score fallback, single-source fallback, partial-failure tolerated, total-failure rethrows) + one self-contained real-gbrain smoke. tests/brain-mcp-tools.test.ts updated for the new kb_search contract (returns gathered hits; fans out per source). Full suite: 1024 pass / 2 fail, both pre-existing env flakes (embedding-gateway env present in this pod; OAuth loopback fake-IdP) — confirmed failing on pristine main.