Executed source 70cf350d23ed3a47ca843754220e4f45fa51aa47: formatting, all six probe tests, real CLI enumeration and strict example Clippy pass on Windows gnullvm. The first five-pass/one-fail run at 05b1b1d is retained. Its declared segment index was independently reconciled against the original saved geometry; the entire successor code change is [2] to [1] in that expectation. No scoring, other expected values or production implementation changed.

Actual f32 enumeration confirms 184 candidates, zero node-clear/stroke-clear choices, and six best (0,0,1) choices. Every best box is 18 world units from its own route and intersects the same foreign ordinary allocation at segment index1. The parent reconciled the complete CLI Project/input bytes and all six owner-tagged winner rows. A second CLI call against its already-existing output fails with AlreadyExists; both input and prior output remain byte-exact. Raw output, gate logs, initial failure, author reconciliation, independent source/arithmetic review, source manifest and parent result remain in this packet.

This rejects candidate A only as specified: ownership-aware tie-breaking inside the existing generator, preserving the best collision tuple and fixed routes, cannot repair the declared nearest-box ownership metric for this reconstructed view. It does not reject new candidate generators/objectives or automatic layout generally, and does not establish that alternate heights cannot aid perception. The exhaustive probe omits production budgets/early exit; no chooser latency or winner simulation is claimed. Restrictions include flat nodes, an automatic unsuspended nonempty target label, no association-end/earlier nonempty labels, and finite orthogonal paths. The representative view is not recovered user source. Native reading of candidate B remains separate.

The parent's first direct guard harness omitted the DLL search path and failed its expected-error assertion before capturing a result. A separately retained repeat returns Windows 0xc0000135 with empty streams; this is a startup failure, not a create-new guard pass. The corrected harness explicitly supplies the pinned libunwind directory and system PATH. That actual repeat reaches AlreadyExists.

From this checkout with Rust configured:

cargo fmt --all -- --check
cargo test --locked -p oa-core --example diagram_readability_probe
cargo clippy --locked -p oa-core --example diagram_readability_probe -- -D warnings
cargo run --locked -p oa-core --example diagram_readability_probe -- label-candidates harness/experiments/diagram-readability-20261001/native-reading/working/eight-card.json examples/release1/expected-diagnostics.json 00000000-0038-4000-8000-000000000001 63009b5f-539e-5914-b630-ec6b7fefb4d5 NEW_OUTPUT_JSON

Choose a nonexistent output path. The collector retains historical absolute paths and performs a deliberate existing-output guard attempt; it is provenance, not a portable replay script. Inventory hashes cover all subpacket files except the inventory itself. Runtime provenance requires the external manifest/source gates; the JSON's embedded experiment description alone is not a build identity.