The earlier commits on this branch met `max_calls = 4`, which measured
the wrong thing: two thirds of what it flagged was a five-stage pipeline
reading down one line, and splitting those bought a binding named
`declared` and nothing else.
What replaces it, measured against this workspace, is
`max_calls = 5` with a closure carrying more than one line of its own
counted as two stages, plus a rule for the one shape that is always two
computations in an expression: building a collection, walking it again,
and building another.
So this reverts the 64 files whose only chains were five stages on a
line, and finishes the ones the new pair asks for: the four NAPI worker
threads now build, spawn, and report in three statements; pnpr's access
log gets `access_span` and `record_access` instead of two closures
inside a `TraceLayer` chain; and `with_boolean_negations` hands its flag
walk to `negations_to_add`.
What the earlier commits built stays, because it stands without a rule:
the shared `join_all_buffered`, `node_eval_command`, `query_flag`, and
`walk_sorted` helpers, the deleted `hosted_revision_sources` duplicate,
and the resolver's context wiring.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CWntUqQ2mPR1yxZNbAYNUg
