F342 reproducibility and evidence boundary
Date: 2026-08-10 UTC
Platform: Linux 7.0.0-28-generic x86_64
Python: 3.12.3
MPI: Open MPI 4.1.6
Filesystem under test: local overlayfs / temporary directory

Production mechanism:
- external_hashes remains the complete set of stable external content observations
- final accounting streams canonical 64-hex no-follow regular public entries
- the same pass counts public objects and membership in external_hashes
- present external is |public intersection observed-external|
- generated/new is |public| - |public intersection observed-external|
- phantom external observations cannot suppress a public generated object
- metadata, directories, noncanonical names, and canonical-looking links remain excluded
- no listdir or public-name materialization is introduced

Automated tests:
- directed lifecycle: 49 passed, 40 subtests passed, 0.90 s
- related six-module integration: 338 passed, 67 subtests passed, 16.58 s
- full warnings-as-errors Python: 731 passed, 87 subtests passed, 94.03 s
- Ruff over production, test, verifier, MPI driver, and benchmark: passed
- py_compile over the same five Python files: passed
- git diff --check: passed
- SVG XML parse and 1900x1400 browser rendering: passed and visually reviewed
- Codex delivery verifier including F302-F342 recomputation: passed

Actual MPI transport:
- one Open MPI run, 7 ranks, 2 masters, 5 workers, one physical host
- rank 0 observes one external hash whose HRW owner is delayed rank 1
- the evidence driver removes that source after rank-0 observation and before owner service
- one pre-existing public generated object is executed exactly once
- phantom external is never public; final public namespace is the generated hash only
- final external=0 and new=1, returncode=0, active/retired epoch=0/1
- worker shutdown acknowledges 3/3 and 2/2; no staging residue remains
- all 10 structured MPI checks pass

Interleaved mechanism benchmark:
- 4096 canonical empty regular public objects on a local temporary directory
- 2048 present external hashes, 2048 generated hashes, and 2048 phantom external hashes
- five warmups and 30 retained observations per mechanism
- deterministic per-round order shuffle; all 60 raw observations are preserved
- F341 counterfactual: public=4096, generated=0 (incorrect)
- F342 intersection: public=4096, external=2048, generated=2048 (exact)
- legacy count-only median=10383.5935 us, P95=14124.431 us
- streamed-intersection median=10566.924 us, P95=11058.167 us
- median intersection/legacy ratio=1.0176557855428374 (about 1.77% overhead)

Strict boundary:
- The MPI run uses an explicit rank delay and a synthetic observer target.
- The benchmark uses empty, metadata-hot local files and measures final scan cost only.
- The final scan trusts prior content-addressed admission and does not rehash every object.
- external_hashes and imported-name identity state remain O(number of observations).
- No real multi-host, NFS/Lustre/GPFS, power-loss, or storage-failover evidence exists.
- No DSE throughput, solver, coverage, campaign, bug-discovery, or LAVA-M uplift is claimed.