F343 bounded streaming simulation evidence
Date: 2026-08-10 UTC
Repository base commit: 146e01b2d6f8e02fd2526764d46ea0d80a4bb6f6 (dirty research worktree)
Host: Linux 7.0.0-28-generic x86_64 GNU/Linux
Python: 3.12.3
MPI: Open MPI 4.1.6
Evidence filesystem: overlay

Implementation
- result object and logical-byte budgets are proven before output creation
- source is one O_NOFOLLOW regular fd with before/after/path identity closure
- RNG draw order and repeated-position last-write-wins behavior are preserved
- source is read in 1 MiB chunks and outputs in batches of at most 32 fds
- only the currently patched chunk is materialized as a bytearray
- partial writes are completed by a shared descriptor helper
- all outputs remain private temporary names until source verification
- hard-link publication is atomic/no-clobber; partial publication is rolled back
- run_symcc passes its authoritative result budgets to simulation generation

Directed regression
- 53 passed + 43 subtests, warnings as errors, 0.92 s
- fixed-seed legacy byte equivalence across 31-byte chunks and 2-output batches
- forced <=7-byte partial writes complete correctly
- object/byte budgets reject before output directory creation
- empty/symlink source, pre-existing target, and publish-time racer covered
- write failure, source identity change, and second-link failure leave no output
- 8 MiB input / 5 outputs / 64 KiB chunk traced peak remains below 2 MiB

Related regression
- 342 passed + 70 subtests, warnings as errors, 16.79 s

Full Python regression
- 735 passed + 90 subtests, warnings as errors, 95.02 s

Mechanism benchmark
- 2 warm-ups + 10 retained samples per mechanism and scale
- 8 MiB legacy/new traced peak: 25,175,994 / 3,152,950 bytes (7.984901x)
- 32 MiB legacy/new traced peak: 100,673,466 / 3,152,990 bytes (31.929523x)
- 8 MiB legacy/new RSS median: 66,968 / 40,098 KiB (1.670108x)
- 32 MiB legacy/new RSS median: 133,868 / 39,866 KiB (3.357949x)
- 8 MiB legacy/new elapsed median: 22,788.8745 / 10,617.483 us (2.146354x)
- 32 MiB legacy/new elapsed median: 138,963.6725 / 28,821.772 us (4.821483x)
- 4x input growth changes legacy/new traced peak by 3.998788x / 1.000013x
- every retained count, total-byte result, and ordered output vector is exact

Real Open MPI
- actual mpirun -np 2: 1 master + 1 worker, production --simulate path
- return 0 in 2.081313 s; result budget 5 objects / 5 bytes is reported
- 1 external seed + 5 generated public one-byte objects
- 6/6 public objects executed exactly once; 10 generated observations
- 1/1 worker ACK; active/retired epochs 0/1; staging residue 0
- 13/13 structured checks true

Claim boundary
- mechanism and same-host integration evidence only
- synthetic mutation/observer target; no symbolic solver invocation
- no multi-host/NFS/Lustre, campaign, coverage, bug-discovery, or LAVA-M claim
- tracemalloc excludes native/kernel allocations; ru_maxrss is fresh-process peak
- local timing describes this Linux/overlayfs host and is not campaign speedup