Fuzz campaign records

Completed AFL++ campaigns, newest first. Each record states the revision, host, duration and outcome of one bounded run. They are historical evidence for compatibility and validation, not a guarantee for later revisions. To run a campaign yourself, follow the AFL guide; fuzzing mechanics describes the targets and oracles.

Interrupted AFL campaign — 2026-10-09

After the distance-cache prefilter (q7–q11), the u8x32 dictionary compare and the UTF-8 early exit (q10/q11), fresh builds of the uncommitted working tree on v0.5.6 ran campaign.sh with CAMPAIGN_PARALLEL=1 (24 stable and 26 experimental compression-side targets) and decoder-campaign.sh (7 stable and 8 experimental decoder targets): 65 AFL++ workers on the i7-13700KF/WSL2 host’s 24 threads, fixed 30 s and 5 s timeouts. The run was planned for 4.5 hours and stopped by hand after 1 h 40 min (5,988–6,066 s per worker) to free the host for benchmarks.

No crashes, hangs or correctness-oracle failures were recorded; stability stayed at or above 99.39%. The changed boundaries are slow under instrumentation on this oversubscribed host: q11_roundtrip and q10_roundtrip reached 10,067–14,330 executions, encoder_session 16,578, so this run is shallow for q10/q11. The committed regression corpus replayed cleanly through cargo afl test in both builds before the campaign.

Four-hour AFL campaign — 2026-10-01

After the greedy-encoder changes of 2026-10-01 (slice-fed range stores for the deep q7–q9 bucket shapes, a flat position and tag index for q5–q6, slice-fed literal splitting in the meta-block builder), fresh builds of the working tree ran for 4 hours: campaign.sh with CAMPAIGN_PARALLEL=1 (24 stable and 26 experimental compression-side targets) and decoder-campaign.sh (7 stable and 8 experimental decoder targets), 65 AFL++ workers on the i7-13700KF/WSL2 host’s 24 threads, with the scripts’ fixed 30 s and 5 s timeouts. Every worker ran its full 14,400 s.

No crashes or correctness-oracle failures were recorded; stability stayed at or above 99.39%. encoder_session (stable) saved five hangs, none of them a loop or touching the changed code. Four are quality 11 over 128 KiB in 4 KiB chunks: one finishes in 12.7 s of CPU time standalone under instrumentation, 35.5 s of wall time on the oversubscribed host, above the 30 s timeout. The fifth is the quality 10, one-byte-chunk case of the 2026-09-27 campaign: 125 s of CPU time standalone, again search cost scaling with payload length. Both inputs exit cleanly. The committed regression corpus replayed cleanly through cargo afl test in both builds before the campaign.

Decoder command-loop smoke campaign — 2026-09-30

After the command-loop changes of 2026-09-30 (slice-cursor refill, per-block-type distance tables, snapshot copies tried before the exact-extent test, one combined copy-limit test, no per-write ring-end test for linear history), decoder-campaign.sh ran every decoder target of both builds for 20 minutes with its fixed 5 s timeout: 15 target/build pairs, 22,352,094 stable and 29,512,959 experimental executions (51,865,053 in total). No crashes or hangs were recorded; stability stayed at or above 99.39%. The committed regression corpus replays cleanly through cargo afl test in both builds.

Eight-hour AFL campaign and follow-up — 2026-09-27/28

After uninitialized output landed (17b3d315), every fuzz target of both feature builds ran for 8 hours on 12 threads of the i7-13700KF/WSL2 host: 66 target/build pairs (30 stable, 36 experimental, compression and decompression) scheduled through twelve slots. The 14 pairs whose oracles reach uninitialized output (encoder_session, decoder_session, streaming_equivalence, output_capacity, decompress, decode_streaming, c_abi) ran 2 hours each and the others 75 minutes, with the fixed timeouts of campaign.sh (30 s) and decoder-campaign.sh (5 s).

No crashes or correctness-oracle failures were recorded; stability stayed at or above 99.38%. encoder_session saved one hang per build. Both inputs finish in about 120 s standalone under instrumentation, and the time scales linearly with payload length (14, 31, 60 and 119 s for an eighth, a quarter, a half and all of it): quality 10 with one-byte chunks and a flush every three blocks, over roughly ten compressions of a 125 KiB payload. That is search cost, not a loop. AFL++ 4.40 ended four workers at start-up with an expired -V limit (reporting “total runtime infty days”); each was rerun for its full slot.

Performance fixes to the decoder and the C ABI followed, and the twelve decoder and C ABI target/build pairs they touch ran for 30 minutes more on 12 threads (36,517,398 executions). decompress saved nine crashes, all one bug that predates uninitialized output: a session growing its history ring for the whole of a long literal insert failed a workspace limit that decompress_to_slice into the same buffer never meets. The minimized input is fuzz/afl/regressions/decompress/crash-long-insert-ring-growth.bin and the a_long_insert_grows_the_ring_only_for_output_the_call_can_hold test; both failed before the fix, and all nine crashes pass after it. On the final tree, decompress, decoder_session, decode_streaming and c_abi in both builds, decode_io_limits and decode_lifecycle in the stable build, and framed_decode and decode_serialized in the experimental one ran 15 minutes more (25,065,107 executions) with no crashes or hangs.

Six-hour AFL campaign — 2026-09-26

After the decoder changes of 2026-09-26 (unmasked refill, literal tail runs, linear slice history), every fuzz target ran for 6 hours from fresh builds of the working tree: campaign.sh with CAMPAIGN_PARALLEL=1 (22 stable and 24 experimental encoder-side targets), decoder-campaign.sh (6 stable and 7 experimental decoder targets) and the three framed targets that neither script covers (framed_decode, framed_encode, framed_roundtrip, experimental build), for 62 AFL++ workers on the i7-13700KF/WSL2 host’s 24 threads.

No crashes or correctness-oracle failures were recorded; stability stayed at or above 99.38%. The decompress target now also checks that decompress_to_slice matches ring-backed session delivery, byte for byte.

framed_decode saved two hangs at its 5 s timeout. Both inputs finish in 0.57 s standalone under instrumentation, so the timeouts came from oversubscription, not an unbounded loop. Their time goes to framing::core::Engine::framing_bytes: it recomputes workspace accounting over every chunk and metadata record on each budget check, which is quadratic in the number of records. This campaign found that slow path in the experimental framed decoder; it is open, and it is not a decoding error.

Eight-hour AFL campaign — 2026-09-11

The recorded campaign exercised compression and decompression in stable and experimental builds for 8 hours, using 59 AFL++ workers on an Intel Core i7-13700KF/WSL2 host. Across 914,882,807 executions, no crashes or correctness-oracle failures were recorded. Checks included C compatibility, round trips, API/chunking/backend equivalence, dictionaries, limits and lifecycle.

The run found one allocation-related slowdown in large_window (19 saved hangs). It was fixed and covered by hq_forest_allocation. Both affected feature builds were then fuzzed for 45 minutes each, starting with the saved inputs: 18,400 further executions, no crashes or hangs. This is a completed validation record, not a guarantee for all possible inputs or later revisions.