v3.11.6-rc.15 (audit H-1) — how many fan-out sub-queries run concurrently.
Bounds the simultaneous full pipelines (graph-boost / optional reranker /
embedding) so a legal 9-phrasing request can't start 9 heavy pipelines at
once on the event loop. Combined with the buildTfidfIndex single-flight
(which collapses the shared corpus BUILD — the expensive read+tokenize scan,
the H-1 amplifier — to ONE regardless), this bounds the worst-case cold
fan-out cost. (rc.16 precision: the corpus build is now vault-size-
independent under fan-out; each sub-query still runs its own
vault.listMarkdown() directory walk to compute the corpus snapshot — cheap
threadpool readdir/stat metadata I/O, not the content read that made H-1
HIGH — so total cold cost is dominated by the single build, not the N walks.)
v3.11.6-rc.15 (audit H-1) — how many fan-out sub-queries run concurrently. Bounds the simultaneous full pipelines (graph-boost / optional reranker / embedding) so a legal 9-phrasing request can't start 9 heavy pipelines at once on the event loop. Combined with the
buildTfidfIndexsingle-flight (which collapses the shared corpus BUILD — the expensive read+tokenize scan, the H-1 amplifier — to ONE regardless), this bounds the worst-case cold fan-out cost. (rc.16 precision: the corpus build is now vault-size- independent under fan-out; each sub-query still runs its ownvault.listMarkdown()directory walk to compute the corpus snapshot — cheap threadpool readdir/stat metadata I/O, not the content read that made H-1 HIGH — so total cold cost is dominated by the single build, not the N walks.)