(fix) hybrid scoring: compute the score max in a loop — the Math.max(...) spread crashed at 130K+ docs

[ranxianglei]

Problem

hybridAlgorithm.score — the default search algorithm — crashes with RangeError: Maximum call stack size exceeded when the doc count gets large. Reproduced with 160,170 docs (46MB corpus, the scale of a long compressed session): the call throws before returning any result, so search breaks entirely instead of degrading. Found while analyzing Tyan66666/billion-context-dsh#133 (search latency in long sessions); reported as #227.

Cause

src/search/algorithms/hybrid.ts normalizes the two channel scores with an argument spread:

const maxBm = Math.max(...bm.map((r) => r.score), 1e-9)

Math.max(...array) passes the whole array as call arguments, and V8 caps call arguments at ~125K–130K (Node 22: 125,000 works, 130,000 throws). Any corpus above that limit — ~30K+ compressed messages in a long session — crashes the search.

Fix

Compute the max with a loop (maxScore helper) — same semantics (max(floor, ...scores)), no argument limit, O(n) either way. Two call sites changed; scoring behavior is byte-identical below the limit.

Verification

• New regression test tests/hybrid-large-corpus.test.ts: 130,000 docs (just above the Node 22 limit) — the old code fails it with RangeError (verified by stashing the fix), the fixed code passes in ~1.1s and still ranks the distinctive doc first (normalization intact).
• Full suite: 625/625 pass; tsc --noEmit clean.

[github-actions]

📦 Built Package Artifact

Branch: 2026-09-09_hybrid-max-arg-limit (9b5f022)

Option A — Install from npm PR tag (recommended)

In your adapter project:

npm install acp-kernel@pr-228

Each push to this PR publishes a new version under the pr-228 npm tag.

Option B — Download artifact

1. Download the artifact from the Actions run
2. Extract the tarball and install:

tar xzf acp-kernel-pr228.tgz
npm install ./package

---

This comment is automatically updated on each push.