Modernization Phase 2.4: streamr-dht modules + bench checkpoint — pivot to coarse façade partitions [iosbuild]

[ptesavol]

The major bench checkpoint did exactly what it exists for: it caught an architecture flaw by measurement, and the fix turned a 2.3× regression into the migration's first real win.

The checkpoint story

The plan's original façade design — one partition per public header — multiplied compile costs at dht scale: 45 partitions each re-parsed the folly/protobuf header stack in their global module fragments, and the BMI dependency chain amplified incremental cascades (every partition BMI of logger→utils→proto-rpc→dht rebuilt in sequence before any test TU).

Metric (macOS, idle, root tree)	2.0 baseline	per-header partitions	coarse partitions (this PR)
Clean build	102 s	238 s ❌	83 s — −19 % ✅
SLogger.hpp incremental	62–70 s	336 s ❌	59 s ✅
StreamID.hpp incremental	19 s	54 s	28 s (BMI-chain latency; honest small regression)

Design revision (all packages, in this PR): each package exposes ONE :all façade partition (+ :protos where generated code exists). The import streamr.<pkg>; interface is unchanged; per-header granularity was only ever needed for consolidation-stage code placement and returns there. Clean build is already 19 % under the pre-modules baseline at the façade stage — before any consolidation gains — because the ~100 test TUs now load BMIs instead of re-parsing header stacks.

streamr-dht migration (the payload)

• :protos wraps the DhtRpc trio — 35 messages, 6 enums (using enum for enumerators), 9+9 generated client/server stubs. The 2.3 protobuf overlay patch held for the 12k-line header.
• :all over all 46 public headers, 6 nested event namespaces re-exported.
• 43 test TUs flipped to import streamr.dht — the flip needed exactly three one-line fixes.
• The plan's headline risk (connection/endpoint include-cycle → partition DAG) was a non-issue at the façade stage, as partitions don't reference each other; the question moves to consolidation as designed.

Two more durable findings

1. Host-environment hazard removed: homebrewClang.cmake injected $HOMEBREW_PREFIX/include as standard include dirs; exported module targets baked them in, and consumer-side synthesized BMI compiles ordered them FIRST — a brew-installed folly/fmt silently shadowed vcpkg's. The global include dirs are gone (keg clang finds its own headers; all deps come from vcpkg). Latent footgun for any machine with brew C++ libraries installed, exposed only by the modules export machinery.
2. Rule hardened: import-using TUs must not also textually include sibling-package headers reaching the same third-party stacks — clangd's experimental modules support reports false folly-ODR violations when its preamble and BMI views overlap. All remaining sibling includes in tests flipped to imports; zero new lint exclusions.

Verification

Gate	Result
dht 81/81, utils 49/49 (re-run), all suites via import	✅
tn + proxyclient downstream builds	✅ unchanged
Root tree	✅ 307/307
Full lint	✅
Bench checkpoint vs 2.0 baselines	recorded in MODERNIZATION.md
iOS cross-build	this PR ([iosbuild])

Next

Phase 2.5: trackerless-network + proxyclient (final metrics), then the 2.6 consolidations.

🤖 Generated with Claude Code